Medical Image Computing and Computer-Assisted Intervention – MICCAI 2014

The large size of histological images combined with their very challenging appearance are two main difficulties which considerably complicate their analysis. In this paper, we introduce an interactive strategy leveraging the output of a supervised random forest classifier to guide a user through such large visual data. Starting from a forest-based pixelwise estimate, subregions of the images at hand are automatically ranked and sequentially displayed according to their expected interest. After each region suggestion, the user selects among several options a rough estimate of the true amount of foreground pixels in this region. From these one-click inputs, the region scoring function is updated in real time using an online gradient descent procedure, which corrects on-the-fly the shortcomings of the initial model and adapts future suggestions accordingly. Experimental validation is conducted for extramedullary hematopoesis localization and demonstrates the practical feasibility of the procedure as well as the benefit of the online adaptation strategy.

Cell detection and segmentation in microscopy images is important for quantitative high-throughput experiments. We present a learning-based method that is applicable to different modalities and cell types, in particular to cells that appear almost transparent in the images. We first train a classifier to detect (partial) cell boundaries. The resulting predictions are used to obtain superpixels and a weighted region adjacency graph. Here, edge weights can be either positive (attractive) or negative (repulsive). The graph partitioning problem is then solved using correlation clustering segmentation. One variant we newly propose here uses a length constraint that achieves state-of-art performance and improvements in some datasets. This constraint is approximated using non-planar correlation clustering. We demonstrate very good performance in various bright field and phase contrast microscopy experiments.

The automatic reconstruction of neurons from stacks of electron microscopy sections is an important computer vision problem in neuroscience. Recent advances are based on a two step approach: First, a set of possible 2D neuron candidates is generated for each section independently based on membrane predictions of a local classifier. Second, the candidates of all sections of the stack are fed to a neuron tracker that selects and connects them in 3D to yield a reconstruction. The accuracy of the result is currently limited by the quality of the generated candidates. In this paper, we propose to replace the heuristic set of candidates used in previous methods with samples drawn from a conditional random field (CRF) that is trained to label sections of neural tissue. We show on a stack of

Drosophila melanogaster

neural tissue that neuron candidates generated with our method produce 30% less reconstruction errors than current candidate generation methods. Two properties of our CRF are crucial for the accuracy and applicability of our method: (1) The CRF models the orientation of membranes to produce more plausible neuron candidates. (2) The interactions in the CRF are restricted to form a bipartite graph, which allows a great sampling speed-up without loss of accuracy.

Models of the diffusion-weighted signal are of strong interest for population studies of the brain microstructure. These studies are typically conducted by extracting a scalar property from the model and subjecting it to null hypothesis significance testing. This process has two major limitations: the reported p-value is a weak predictor of the reproducibility of findings and evidence for the absence of microstructural alterations cannot be gained. To overcome these limitations, this paper proposes a Bayesian framework for population studies of the brain microstructure represented by multi-fascicle models. A hierarchical model is built over the biophysical parameters of the microstructure. Bayesian inference is performed by Hamiltonian Monte Carlo sampling and results in a joint posterior distribution over the latent microstructure parameters for each group. Inference from this posterior enables richer analyses of the brain microstructure beyond the dichotomy of significance testing. Using synthetic and in-vivo data, we show that our Bayesian approach increases reproducibility of findings from population studies and opens new opportunities in the analysis of the brain microstructure.

Many microscopic imaging modalities suffer from the problem of intensity inhomogeneity due to uneven illumination or camera nonlinearity, known as shading artifacts. A typical example of this is the unwanted seam when stitching images to obtain a whole slide image (WSI). Elimination of shading plays an essential role for subsequent image processing such as segmentation, registration, or tracking. In this paper, we propose two new retrospective shading correction algorithms for WSI targeted to two common forms of WSI: multiple image tiles before mosaicking and an already-stitched image. Both methods leverage on recent achievements in matrix rank minimization and sparse signal recovery. We show how the classic shading problem in microscopy can be reformulated as a decomposition problem of low-rank and sparse components, which seeks an optimal separation of the foreground objects of interest and the background illumination field. Additionally, a sparse constraint is introduced in the Fourier domain to ensure the smoothness of the recovered background. Extensive qualitative and quantitative validation on both synthetic and real microscopy images demonstrates superior performance of the proposed methods in shading removal in comparison with a well-established method in ImageJ.

We propose a novel cell segmentation approach by estimating a cell-sensitive camera response function based on variously exposed phase contrast microscopy images on the same cell dish. Using the cell-sensitive microscopy imaging, cells’ original irradiance signals are restored from all exposures and the irradiance signals on non-cell background regions are restored as a uniform constant (i.e., the imaging system is sensitive to cells only but insensitive to non-cell background). Cell segmentation is then performed on the restored irradiance signals by simple thresholding. The experimental results validate that high quality cell segmentation can be achieved by our approach.

We describe a technique that employs the stochastic Latent Topic Models framework to allow quantification of melanin and hemoglobin content in dermoscopy images. Such information bears useful implications for analysis of skin hyperpigmentation, and for classification of skin diseases. The proposed method outperforms existing approaches while allowing for more stringent and probabilistic modeling than previously.

Stem cell-derived cardiomyocytes hold tremendous potential for drug development and safety testing related to cardiovascular health. The characterization of cardiomyocytes is most commonly performed using electrophysiological systems, which are expensive, laborious to use, and may induce undesirable cellular response. Here, we present a new method for non-invasive characterization of cardiomyocytes using video microscopy and image analysis. We describe an automated pipeline that consists of segmentation of beating regions, robust beating signal calculation, signal quantification and modeling, and hierarchical clustering. Unlike previous imaging-based methods, our approach enables clinical applications by capturing beating patterns and arrhythmias across healthy and diseased cells with varied densities. We demonstrate the strengths of our algorithm by characterizing the effects of two commercial drugs known to modulate beating frequency and irregularity. Our results provide, to our knowledge, the first clinically-relevant demonstration of a fully-automated and non-invasive imaging-based beating assay for characterization of stem cell-derived cardiomyocytes.

In this paper, we improve upon earlier approaches to segmenting mitochondria in Electron Microscopy images by explicitly modeling the double membrane that encloses mitochondria, as well as using features that capture context over an extended neighborhood. We demonstrate that this results in both improved classification accuracy and reduced computational requirements for training.

Identifying neutrophils lays a crucial foundation for diagnosing acute inflammation diseases. But, such computerized methods on the commonly used H&E staining histology tissue images are lacking, due to various inherent difficulties of identifying cells in such image modality and the challenge that a considerable portion of neutrophils do not have a “textbook” appearance. In this paper, we propose a new method for identifying neutrophils in H&E staining histology tissue images. We first segment the cells by applying iterative edge labeling, and then identify neutrophils based on the segmentation results by considering the “context” of each candidate cell constructed by a new Voronoi diagram of clusters of other neutrophils. We obtain good performance compared with two baseline algorithms we constructed, on clinical images collected from patients suspected of having inflammatory bowl diseases.

In this work we present a novel technique we term

active graph matching

, which integrates the popular active shape model into a sparse graph matching problem. This way we are able to combine the benefits of a global, statistical deformation model with the benefits of a local deformation model in form of a second-order random field. We present a new iterative energy minimization technique which achieves empirically good results. This enables us to exceed state-of-the art results for the task of annotating nuclei in 3D microscopic images of

C. elegans

. Furthermore with the help of the generalized Hough transform we are able to jointly segment and annotate a large set of nuclei in a fully automatic fashion for the first time.

Content-based video retrieval has shown promising results to help physicians in their interpretation of medical videos in general and endomicroscopic ones in particular. Defining a relevant query for CBVR can however be a complex and time-consuming task for non-expert and even expert users. Indeed, uncut endomicroscopy videos may very well contain images corresponding to a variety of different tissue types. Using such uncut videos as queries may lead to drastic performance degradations for the system. In this study, we propose a semi-automated methodology that allows the physician to create meaningful and relevant queries in a simple and efficient manner. We believe that this will lead to more reproducible and more consistent results. The validation of our method is divided into two approaches. The first one is an indirect validation based on per video classification results with histopathological ground-truth. The second one is more direct and relies on perceived inter-video visual similarity ground-truth. We demonstrate that our proposed method significantly outperforms the approach with uncut videos and approaches the performance of a tedious manual query construction by an expert. Finally, we show that the similarity perceived between videos by experts is significantly correlated with the inter-video similarity distance computed by our retrieval system.

We present a new algorithm for automatic and interactive segmentation of neuron structures from electron microscopy (EM) images. Our method selects a collection of nodes from the watershed merging tree as the proposed segmentation. This is achieved by building a conditional random field (CRF) whose underlying graph is the merging tree. The maximum a posteriori (MAP) prediction of the CRF is the output segmentation. Our algorithm outperforms state-of-the-art methods. Both the inference and the training are very efficient as the graph is tree-structured. Furthermore, we develop an interactive segmentation framework which selects uncertain regions for a user to proofread. The uncertainty is measured by the marginals of the graphical model. Based on user corrections, our framework modifies the merging tree and thus improves the segmentation globally.

Medical image segmentation has traditionally been regarded as a separate process from image acquisition and reconstruction, even though its performance directly depends on the quality and characteristics of these first stages of the imaging pipeline. Adopting an integrated acquisition-reconstruction-segmentation process can provide a more efficient and accurate solution. In this paper we propose a joint segmentation and reconstruction algorithm for undersampled magnetic resonance data. Merging a reconstructive patch-based sparse modelling and a discriminative Gaussian mixture modelling can produce images with enhanced edge information ultimately improving their segmentation.

In this paper we propose a novel algorithm for jointly performing data based motion correction and direct parametric reconstruction of dynamic PET data. We derive a closed form update for the penalised likelihood maximisation which greatly enhances the algorithm’s computational efficiency for practical use. Our algorithm achieves sub-voxel motion correction residual with noisy data in the simulation-based validation and reduces the bias of the direct estimation of the kinetic parameter of interest. A preliminary evaluation on clinical brain data using [

18

F]Choline shows improved contrast for regions of high activity. The proposed method is based on a data-driven kinetic modelling method and is directly applicable to reversible and irreversible PET tracers, covering a range of clinical applications.

The reconstruction of a 3D volume from a stack of 2D histology slices is still a challenging problem especially if no external references are available. Without a reference, standard registration approaches tend to align structures that should not be perfectly aligned. In this work we introduce a deformable, reference-free reconstruction method that uses an internal structural probability map (SPM) to regularize a free-form deformation. The SPM gives an estimate of the original 3D structure of the sample from the misaligned and possibly corrupted 2D slices. We present a consecutive as well as a simultaneous reconstruction approach that incorporates this estimate in a deformable registration framework. Experiments on synthetic and mouse brain datasets indicate that our method produces similar results compared to reference-based techniques on synthetic datasets. Moreover, it improves the smoothness of the reconstruction compared to standard registration techniques on real data.

Low-resolution and signal-dependent noise distribution in positron emission tomography (PET) images makes denoising process an inevitable step prior to qualitative and quantitative image analysis tasks. Conventional PET denoising methods either over-smooth small-sized structures due to resolution limitation or make incorrect assumptions about the noise characteristics. Therefore, clinically important quantitative information may be corrupted. To address these challenges, we introduced a novel approach to remove signal-dependent noise in the PET images where the noise distribution was considered as

Poisson-Gaussian

mixed. Meanwhile, the generalized Anscombe’s transformation (GAT) was used to stabilize varying nature of the PET noise. Other than noise stabilization, it is also desirable for the noise removal filter to preserve the boundaries of the structures while smoothing the noisy regions. Indeed, it is important to avoid significant loss of quantitative information such as standard uptake value (SUV)-based metrics as well as metabolic lesion volume. To satisfy all these properties, we extended bilateral filtering method into trilateral filtering through multiscaling and optimal Gaussianization process. The proposed method was tested on more than 50 PET-CT images from various patients having different cancers and achieved the superior performance compared to the widely used denoising techniques in the literature.

In this study, we propose a novel scheme for real time dynamic magnetic resonance imaging (dMRI) reconstruction. Different from previous methods, the reconstructions of the second frame to the last frame are independent in our scheme, which only require the first frame as the reference. Therefore, this scheme can be naturally implemented in parallel. After the first frame is reconstructed, all the later frames can be processed as soon as the

k

-space data is acquired. As an extension of the convention total variation, a new online model called dynamic total variation is used to exploit the sparsity on both spatial and temporal domains. In addition, we design an accelerated reweighted least squares algorithm to solve the challenging reconstruction problem. This algorithm is motivated by the special structure of partial Fourier transform in sparse MRI. The proposed method is compared with 4 state-of-the-art online and offline methods on in-vivo cardiac dMRI datasets. The results show that our method significantly outperforms previous online methods, and is comparable to the offline methods in terms of reconstruction accuracy.

The reconstruction of 4D images from 2D navigator and data slices requires sufficient observations per motion state to avoid blurred images and motion artifacts between slices. Especially images from rare motion states, like deep inhalations during free-breathing, suffer from too few observations.

To address this problem, we propose to actively generate more suitable images instead of only selecting from the available images. The method is based on learning the relationship between navigator and data-slice motion by linear regression after dimensionality reduction. This can then be used to predict new data slices for a given navigator by warping existing data slices by their predicted displacement field. The method was evaluated for 4D-MRIs of the liver under free-breathing, where sliding boundaries pose an additional challenge for image registration.

Leave-one-out tests for five short sequences of ten volunteers showed that the proposed prediction method improved on average the residual mean (95%) motion between the ground truth and predicted data slice from 0.9mm (1.9mm) to 0.8mm (1.6mm) in comparison to the best selection method. The approach was particularly suited for unusual motion states, where the mean error was reduced by 40% (2.2mm vs. 1.3mm).

Acute brain diseases such as acute stroke and transit ischemic attacks are the leading causes of mortality and morbidity worldwide, responsible for 9% of total death every year. ‘Time is brain’ is a widely accepted concept in acute cerebrovascular disease treatment. Efficient and accurate computational framework for hemodynamic parameters estimation can save critical time for thrombolytic therapy. Meanwhile the high level of accumulated radiation dosage due to continuous image acquisition in CT perfusion (CTP) raised concerns on patient safety and public health. However, low-radiation will lead to increased noise and artifacts which require more sophisticated and time-consuming algorithms for robust estimation. We propose a novel efficient framework using tensor total-variation (TTV) regularization to achieve both high efficiency and accuracy in deconvolution for low-dose CTP. The method reduces the necessary radiation dose to only 8% of the original level and outperforms the state-of-art algorithms with estimation error reduced by 40%. It also corrects over-estimation of cerebral blood flow (CBF) and under-estimation of mean transit time (MTT), at both normal and reduced sampling rate. An efficient computational algorithm is proposed to find the solution with fast convergence.

Speckle noise is problematic in optical coherence tomography (OCT). With the fast scan rate, swept source OCT scans the same position in the retina for multiple times rapidly and computes an average image from the multiple scans for speckle reduction. However, the eye movement poses some challenges. In this paper, we propose a new method for speckle reduction from multiply-scanned OCT slices. The proposed method applies a preliminary speckle reduction on the OCT slices and then registers them using a global alignment followed by a local alignment based on fast iterative diamond search. After that, low rank matrix completion using bilateral random projection is utilized to iteratively estimate the noise and recover the underlying clean image. Experimental results show that the proposed method achieves average contrast to noise ratio 15.65, better than 13.78 by the baseline method used currently in swept source OCT devices. The technology can be embedded into current OCT machines to enhance the image quality for subsequent analysis.

Grating-based X-ray dark-field imaging is a new imaging modality. It allows the visualization of structures at micrometer scale due to small-angle scattering of the X-ray beam. However, reading dark-field images is challenging as absorption and edge-diffraction effects also contribute to the dark-field signal, without adding diagnostic value. In this paper, we present a novel – and to our knowledge the first – algorithm for isolating small-angle scattering in dark-field images, which greatly improves their interpretability. To this end, our algorithm utilizes the information available from the absorption and differential phase images to identify clinically irrelevant contributions to the dark-field image. Experimental results on phantom and ex-vivo breast data promise a greatly enhanced diagnostic value of dark-field images.

A new algorithm for model based registration is presented that optimizes both position and surface normal information of the shapes being registered. This algorithm extends the popular Iterative Closest Point (ICP) algorithm by incorporating the surface orientation at each point into both the correspondence and registration phases of the algorithm. For the correspondence phase an efficient search strategy is derived which computes the most probable correspondences considering both position and orientation differences in the match. For the registration phase an efficient, closed-form solution provides the maximum likelihood rigid body alignment between the oriented point matches. Experiments by simulation using human femur data demonstrate that the proposed Iterative Most Likely Oriented Point (IMLOP) algorithm has a strong accuracy advantage over ICP and has increased ability to robustly identify a successful registration result.

Correspondence matching between MR brain images is often challenging due to large inter-subject structural variability. In this paper, we propose a novel landmark detection method for robust establishment of correspondences between subjects. Specifically, we first annotate distinctive landmarks in the training images. Then, we use regression forest to simultaneously learn (1) the optimal set of features to best characterize each landmark and (2) the non-linear mappings from local patch appearances of image points to their displacements towards each landmark. The learned regression forests are used as landmark detectors to predict the locations of these landmarks in new images. Since landmark detection is performed in the entire image domain, our method can cope with large anatomical variations among subjects. We evaluated our method by applying it to MR brain image registration. Experimental results indicate that by combining our method with existing registration method, obvious improvement in registration accuracy can be achieved.

In traditional free-form deformation (FFD) based registration, a B-spline basis function is commonly utilized to build the transformation model. As the B-spline order increases, the corresponding B-spline function becomes smoother. However, the higher-order B-spline has a larger support region, which means higher computational cost. For a given

D

-dimensional

n

th-order B-spline, an

m

th-order B-spline where (

m

 ≤ 

n

) has

$(\frac{m+1}{n+1})^{D}$

times lower computational complexity. Generally, the third-order B-spline is regarded as keeping a good balance between smoothness and computation time. A lower-order function is seldom used to construct the deformation field for registration since it is less smooth. In this research, we investigated whether lower-order B-spline functions can be utilized for efficient registration, by using a novel stochastic perturbation technique in combination with a postponed smoothing technique to higher B-spline order. Experiments were performed with 3D lung and brain scans, demonstrating that the lower-order B-spline FFD in combination with the proposed perturbation and postponed smoothing techniques even results in better accuracy and smoothness than the traditional third-order B-spline registration, while substantially reducing computational costs.

Medical scans are today routinely acquired using multiple sequences or contrast settings, resulting in multispectral data. For the automatic analysis of this data, the evaluation of multispectral similarity is essential. So far, few concepts have been proposed to deal in a principled way with images containing multiple channels. Here, we present a new approach based on a well known statistical technique: canonical correlation analysis (CCA). CCA finds a mapping of two multidimensional variables into two new bases, which best represent the true underlying relations of the signals. In contrast to previously used metrics, it is therefore able to find new correlations based on linear combinations of multiple channels. We extend this concept to efficiently model local canonical correlation (LCCA) between image patches. This novel, more general similarity metric can be applied to images with an arbitrary number of channels. The most important property of LCCA is its invariance to affine transformations of variables. When used on local histograms, LCCA can also deal with multimodal similarity. We demonstrate the performance of our concept on challenging clinical multispectral datasets.

The random walker image registration (RWIR) method is a powerful tool for aligning medical images that also provides useful uncertainty information. However, it is difficult to ensure topology preservation in RWIR, which is an important property in medical image registration as it is often necessary for the anatomical feasibility of an alignment. In this paper, we introduce a technique for determining spatially adaptive regularization weights for RWIR that ensure an anatomically feasible transformation. This technique only increases the run time of the RWIR algorithm by about 10%, and avoids over-smoothing by only increasing regularization in specific image regions. Our results show that our technique ensures topology preservation and improves registration accuracy.

In this work we propose a novel method to correct echo planar imaging (EPI) distortions in diffusion MRI data acquired with reversed phase encoding directions (“blip-up blip-down” acquisitions). The transformation model is symmetric, diffeomorphic and capable of capturing large deformations. It can take advantage of a structural MRI target and include the contribution of diffusion weighted images, in addition to EPI images acquired without diffusion sensitization. The proposed correction significantly outperform existing strategies, assuring anatomically accurate characterization of the orientation, mean diffusivity, and anisotropy of white matter structures in the human brain.

This paper introduces a variational strategy to learn spatially-varying metrics on large groups of images, in the Large Deformation Diffeomorphic Metric Mapping (LDDMM) framework. Spatially-varying metrics we learn not only favor local deformations but also correlated deformations in different image regions and in different directions. In addition, metric parameters can be efficiently estimated using a gradient descent method. We first describe the general strategy and then show how to use it on 3D medical images with reasonable computational ressources. Our method is assessed on the 3D brain images of the LPBA40 dataset. Results are compared with ANTS-SyN and LDDMM with spatially-homogeneous metrics.

We propose a Sparse Bayesian framework for non-rigid registration. Our principled approach is flexible, in that it efficiently finds an optimal, sparse model to represent deformations among any preset, widely overcomplete range of basis functions. It addresses open challenges in state-of-the-art registration, such as the automatic joint estimate of model parameters (

e.g.

noise and regularization levels). We demonstrate the feasibility and performance of our approach on cine MR, tagged MR and 3D US cardiac images, and show state-of-the-art results on benchmark datasets evaluating accuracy of motion and strain.

The fusion of information from different medical imaging techniques plays an important role in data analysis. Despite the many proposed registration algorithms the problem of registering 2D histological images to 3D CT or MR imaging data is still largely unsolved.

In this paper we propose a computationally efficient automatic approach to match 2D histological images to 3D micro Computed Tomography data. The landmark-based approach in combination with a density-driven RANSAC plane-fitting allows efficient localization of the histology images in the 3D data within less than four minutes (single-threaded MATLAB code) with an average accuracy of 0.25 mm for correct and 2.21 mm for mismatched slices. The approach managed to successfully localize 75% of the histology images in our database. The proposed algorithm is an important step towards solving the problem of registering 2D histology sections to 3D data fully automatically.

Accurate and reliable registration of longitudinal spine images is essential for assessment of disease progression and surgical outcome. Implementing a fully automatic and robust registration for clinical use, however, is challenging since standard registration techniques often fail due to poor initial alignment. The main causes of registration failure are the small overlap between scans which focus on different parts of the spine and/or substantial change in shape (e.g. after correction of abnormal curvature) and appearance (e.g. due to surgical implants). To overcome these issues we propose a registration approach which incorporates estimates of vertebrae locations obtained from a learning-based classification method. These location priors are used to initialize the registration and to provide semantic information within the optimization process. Quantitative evaluation on a database of 93 patients with a total of 276 registrations on longitudinal spine CT demonstrate that our registration method significantly reduces the number of failure cases.

Fusion between an endoscopic movie and a CT can aid specifying the tumor target volume for radiotherapy. That requires a deformable pharyngeal surface registration between a 3D endoscope reconstruction and a CT segmentation. In this paper, we propose to use local geometric features for deriving a set of initial correspondences between two surfaces, with which an association graph can be constructed for registration by spectral graph matching. We also define a new similarity measurement to provide a meaningful way for computing inter-surface affinities in the association graph. Our registration method can deal with large non-rigid anatomical deformation, as well as missing data and topology change. We tested the robustness of our method with synthetic deformations and showed registration results on real data.

Intensity-based image registration requires resampling images on a common grid to evaluate the similarity function. The uncertainty of interpolation varies across the image, depending on the location of resampled points relative to the base grid. We propose to perform Bayesian inference with Gaussian processes, where the covariance matrix of the Gaussian process posterior distribution estimates the uncertainty in interpolation. The Gaussian process replaces a single image with a distribution over images that we integrate into a generative model for registration. Marginalization over resampled images leads to a new similarity measure that includes the uncertainty of the interpolation. We demonstrate that our approach increases the registration accuracy and propose an efficient approximation scheme that enables seamless integration with existing registration methods.

Intensity based registration is a challenge when images to be registered have insufficient amount of information in their overlapping region. Especially, in the absence of dominant structures such as strong edges in this region, obtaining a solution that satisfies global structural consistency becomes difficult. In this work, we propose to exploit the vast amount of available information beyond the overlapping region to support the registration process. To this end, a novel global regularization term using Generalized Hough Transform is designed that ensures the global consistency when the local information in the overlap region is insufficient to drive the registration. Using prior data, we learn a parametrization of the target anatomy in Hough space. This parametrization is then used as a regularization for registering the observed partial images without using any prior data. Experiments on synthetic as well as on sample real medical images demonstrate the good performance and potential use of the proposed concept.

2D/3D registration is a well known technique in medical imaging for combining pre-operative volume data with live fluoroscopy. A common issue of this type of algorithms is that out-of-plane parameters are hard to determine. One solution to overcome this issue is the use of X-ray images from two angulations. However, performing in-plane transformation in one image destroys the registration in the other image, particularly if the angulations are smaller than 90 degrees apart. Our main contribution is the automation of a novel registration approach. It handles translation and rotation of a volume in a way that in-plane parameters are kept invariant and independent of the angle offset between both projections in a double-oblique setting. Our approach yields more robust and partially faster registration results, compared to conventional methods, especially in case of object movement. It was successfully tested on clinical data for fusion of transesophageal ultrasound and X-ray.

We present a new method for rigid registration of CT scans in Radon space. The inputs are the two 3D Radon transforms of the CT scans, one densely sampled and the other sparsely sampled. The output is the rigid transformation that best matches them. The algorithm starts by finding the best matching between each direction vector in the sparse transform and the corresponding direction vector in the dense transform. It then solves the system of linear equations derived from the direction vector pairs. Our method can be used to register two CT scans and to register a baseline scan to the patient with reduced-dose scanning without compromising registration accuracy. Our preliminary simulation results on the Shepp-Logan head phantom dataset and a pair of clinical head CT scans indicates that our 3D Radon space rigid registration method performs significantly better than image-based registration for very few scan angles and comparably for densely-sampled scans.

Recently, patch-based label fusion methods have achieved many successes in medical imaging area. After registering atlas images to the target image, the label at each target image point can be subsequently determined by checking the patchwise similarities between the underlying target image patch and all atlas image patches. Apparently, the definition of patchwise similarity is critical in label fusion. However, current methods often simply use entire image patch with fixed patch size throughout the entire label fusion procedure, which could be insufficient to distinguish complex shape/appearance patterns of anatomical structures in medical imaging scenario. In this paper, we address the above limitations at three folds.

First

, we assign each image patch with multiscale feature representations such that both local and semi-local image information can be encoded to increase robustness of measuring patchwise similarity in label fusion.

Second

, since multiple

variable

neighboring structures could present in one image patch, simply computing patchwise similarity based on the entire image patch is not specific to the particular structure of interest under labeling and can be easily misled by the surrounding

variable

structures in the same image patch. Thus, we partition each atlas patch into a set of new label-specific atlas patches according to the existing label information in the atlas images. Then, the new label-specific atlas patches can be more specific and flexible for label fusion than using the entire image patch, since the complex image patch has now been semantically divided into several distinct patterns.

Finally

, in order to correct the possible mis-labeling, we hierarchically improve the label fusion result in a coarse-to-fine manner by iteratively repeating the label fusion procedure with the gradually-reduced patch size. More accurate label fusion results have been achieved by our hierarchical label fusion method with multiscale feature presentations upon label-specific atlas patches.

We present a novel algorithm for the simultaneous segmentation and anatomical labeling of the cerebral vasculature. The method first constructs an overcomplete graph capturing the vasculature. It then selects and labels the subset of edges that most likely represents the true vasculature. Unlike existing approaches that first attempt to obtain a good segmentation and then perform labeling, we jointly optimize for both by simultaneously taking into account the image evidence and the prior knowledge about the geometry and connectivity of the vasculature. This results in an Integer Program (IP), which we solve optimally using a branch-and-cut algorithm. We evaluate our approach on a public dataset of 50 cerebral MRA images, and demonstrate that it compares favorably against state-of-the-art methods.

We study the widespread, but rarely discussed, tendency of atlas-based segmentation to under-segment the organs of interest. Commonly used error measures do not distinguish between under- and over-segmentation, contributing to the problem. We explicitly quantify over- and under-segmentation in several typical examples and present a new hypothesis for the cause. We provide evidence that segmenting only one organ of interest and merging all surrounding structures into one label creates bias towards background in the label estimates suggested by the atlas. We propose a generative model that corrects for this effect by learning the background structures from the data. Inference in the model separates the background into distinct structures and consequently improves the segmentation accuracy. Our experiments demonstrate a clear improvement in several applications.

The detection of abnormal intensities in brain images caused by the presence of pathologies is currently under great scrutiny. Selecting appropriate models for pathological data is of critical importance for an unbiased and biologically plausible model fit, which in itself enables a better understanding of the underlying data and biological processes. Besides, it impacts on one’s ability to extract pathologically meaningful imaging biomarkers. With this aim in mind, this work proposes a fully unsupervised hierarchical model selection framework for neuroimaging data which permits the stratification of different types of abnormal image patterns without prior knowledge about the subject’s pathological status.

Cochlear Implants (CI) are surgically implanted neural prosthetic devices used to treat severe-to-profound hearing loss. Recent studies have suggested that hearing outcomes with CIs are correlated with the location where individual electrodes in the implanted electrode array are placed, but techniques proposed for determining electrode location have been too coarse and labor intensive to permit detailed analysis on large numbers of datasets. In this paper, we present a fully automatic snake-based method for accurately localizing CI electrodes in clinical post-implantation CTs. Our results show that average electrode localization errors with the method are 0.21 millimeters. These results indicate that our method could be used in future large scale studies to analyze the relationship between electrode position and hearing outcome, which potentially could lead to technological advances that improve hearing outcomes with CIs.

We propose a novel segmentation method based on multi-label graph cuts utilizing higher-order potentials to impose shape priors. Each higher-order potential is defined with respect to a candidate shape, and takes a low value if and only if most of the voxels inside the shape are foreground and most of those outside are background. We apply this technique to coronary lumen and plaque segmentation in CT angiography, exploiting the prior knowledge that the vessel walls tend to be tubular, whereas calcified plaques are more likely globular. We use the Hessian analysis to detect the candidate shapes and introduce corresponding higher-order terms into the energy. Since each higher-order term has any effect only when its highly specific condition is met, we can add many of them at possible locations and sizes without severe side effects. We show the effectiveness of the method by testing it on the standardized evaluation framework presented at MICCAI segmentation challenge 2012. The method achieved values comparable to the best in each of the sensitivity and positive predictive value, placing it at the top in average rank.

The automatic segmentation of cardiac magnetic resonance images poses many challenges arising from the large variation between different anatomies, scanners and acquisition protocols. In this paper, we address these challenges with a global graph search method and a novel spectral embedding of the images. Firstly, we propose the use of an approximate graph search approach to initialize patch correspondences between the image to be segmented and a database of labelled atlases. Then, we propose an innovative spectral embedding using a multi-layered graph of the images in order to capture global shape properties. Finally, we estimate the patch correspondences based on a joint spectral representation of the image and atlases. We evaluated the proposed approach using 155 images from the recent MICCAI SATA segmentation challenge and demonstrated that the proposed algorithm significantly outperforms current state-of-the-art methods on both training and test sets.

Bone localization in ultrasound (US) remains challenging despite encouraging advances. Current methods, e.g. local image phase-based feature analysis, showed promising results but remain reliant on delicate parameter selection processes and prone to errors at confounding soft tissue interfaces of similar appearance to bone interfaces. We propose a different approach combining US strain imaging and envelope power detection at each radio-frequency (RF) sample. After initial estimation of strain and envelope power maps, we modify their dynamic ranges into a modified strain map (MSM) and a modified envelope map (MEM) that we subsequently fuse into a single combined map that we show corresponds robustly to actual bone boundaries. Our quantitative results demonstrate a marked reduction in false positive responses at soft tissue interfaces and an increase in bone delineation accuracy. Comparisons to the state-of-the-art on a finite-element-modelling (FEM) phantom and fiducial-based experimental phantom show an average improvement in mean absolute error (MAE) between actual and estimated bone boundaries of 32% and 14%, respectively. We also demonstrate an average reduction in false bone responses of 87% and 56%, respectively. Finally, we qualitatively validate on clinical in vivo data of the human radius and ulna bones, and demonstrate similar improvements to those observed on phantoms.

Selective and iterative method for performance level estimation (SIMPLE) is a multi-atlas segmentation technique that integrates atlas selection and label fusion that has proven effective for radiotherapy planning. Herein, we revisit atlas selection and fusion techniques in the context of segmenting the spleen in metastatic liver cancer patients with possible splenomegaly using clinically acquired computed tomography (CT). We re-derive the SIMPLE algorithm in the context of the statistical literature, and show that the atlas selection criteria rest on newly presented principled likelihood models. We show that SIMPLE performance can be improved by accounting for exogenous information through Bayesian priors (so called context learning). These innovations are integrated with the joint label fusion approach to reduce the impact of correlated errors among selected atlases. In a study of 65 subjects, the spleen was segmented with median Dice similarity coefficient of 0.93 and a mean surface distance error of 2.2 mm.

Patient-specific orthopedic knee surgery planning requires precisely segmenting from 3D CT images multiple knee bones, namely femur, tibia, fibula, and patella, around the knee joint with severe pathologies. In this work, we propose a fully automated, highly precise, and computationally efficient segmentation approach for multiple bones. First, each bone is initially segmented using a model-based marginal space learning framework for pose estimation followed by non-rigid boundary deformation. To recover shape details, we then refine the bone segmentation using graph cut that incorporates the shape priors derived from the initial segmentation. Finally we remove overlap between neighboring bones using multi-layer graph partition. In experiments, we achieve simultaneous segmentation of femur, tibia, patella, and fibula with an overall accuracy of less than 1mm surface-to-surface error in less than 90s on hundreds of 3D CT scans with pathological knee joints.

This study investigates segmentation with a novel invariant compactness constraint. The proposed prior is a high-order

fractional

term, which is not directly amenable to powerful optimizers. We derive first-order

Gateâux

derivative approximations of our compactness term and adopt an iterative trust region paradigm by splitting our problem into constrained sub-problems, each solving the approximation globally via a Lagrangian formulation and a graph cut. We apply our algorithm to the challenging task of abdominal aorta segmentation in 3D MRI volumes, and report quantitative evaluations over 30 subjects, which demonstrate that the results correlate well with independent manual segmentations. We further show the use of our method in several other medical applications and demonstrate that, in comparison to a standard level-set optimization, our algorithm is one order of magnitude faster.

Pixel and superpixel classifiers have become essential tools for EM segmentation algorithms. Training these classifiers remains a major bottleneck primarily due to the requirement of completely annotating the dataset which is tedious, error-prone and costly. In this paper, we propose an interactive learning scheme for the superpixel classifier for EM segmentation. Our algorithm is ‘active semi-supervised’ because it requests the labels of a small number of examples from user and applies label propagation technique to generate these queries. Using only a small set (< 20%) of all datapoints, the proposed algorithm consistently generates a classifier almost as accurate as that estimated from a complete groundtruth. We provide segmentation results on multiple datasets to show the strength of these classifiers.

In this paper we analyze the properties of the well-known segmentation fusion algorithm STAPLE, using a novel inference technique that analytically marginalizes out all model parameters. We demonstrate both theoretically and empirically that when the number of raters is large, or when consensus regions are included in the model, STAPLE devolves into thresholding the average of the input segmentations. We further show that when the number of raters is small, the STAPLE result may not be the optimal segmentation truth estimate, and its model parameter estimates might not reflect the individual raters’ actual segmentation performance. Our experiments indicate that these intrinsic weaknesses are frequently exacerbated by the presence of undesirable global optima and convergence issues. Together these results cast doubt on the soundness and usefulness of typical STAPLE outcomes.

Hilar dissection is an important and delicate stage in partial nephrectomy during which surgeons remove connective tissue surrounding renal vasculature. Potentially serious complications arise when vessels occluded by fat are missed in the endoscopic view and are not appropriately clamped. To aid in vessel discovery, we propose an automatic method to localize and label occluded vasculature. Our segmentation technique is adapted from phase-based video magnification, in which we measure subtle motion from periodic changes in local phase information albeit for labeling rather than magnification. We measure local phase through spatial decomposition of each frame of the endoscopic video using complex wavelet pairs. We then assign segmentation labels based on identifying responses of regions exhibiting temporal local phase changes matching the heart rate frequency. Our method is evaluated with a retrospective study of eight real robot-assisted partial nephrectomies demonstrating utility for surgical guidance that could potentially reduce operation times and complication rates.

The growing use of image-guided minimally-invasive surgical procedures is confronting clinicians and surgical staff with new radiation exposure risks from X-ray imaging devices. The accurate estimation of intra-operative radiation exposure can increase staff awareness of radiation exposure risks and enable the implementation of well-adapted safety measures. The current surgical practice of wearing a single dosimeter at chest level to measure radiation exposure does not provide a sufficiently accurate estimation of radiation absorption throughout the body. In this paper, we propose an approach that combines data from wireless dosimeters with the simulation of radiation propagation in order to provide a global radiation risk map in the area near the X-ray device. We use a multi-camera RGBD system to obtain a 3D point cloud reconstruction of the room. The positions of the table, C-arm and clinician are then used 1) to simulate the propagation of radiation in a real-world setup and 2) to overlay the resulting 3D risk-map onto the scene in an augmented reality manner. By using real-time wireless dosimeters in our system, we can both calibrate the simulation and validate its accuracy at specific locations in real-time. We demonstrate our system in an operating room equipped with a robotised X-ray imaging device and validate the radiation simulation on several X-ray acquisition setups.

Augmented reality for soft tissue laparoscopic surgery is a growing topic of interest in the medical community and has potential application in intra-operative planning and image guidance. Delivery of such systems to the operating room remains complex with theoretical challenges related to tissue deformation and the practical limitations of imaging equipment. Current research in this area generally only solves part of the registration pipeline or relies on fiducials, manual model alignment or assumes that tissue is static. This paper proposes a novel augmented reality framework for intra-operative planning: the approach co-registers pre-operative CT with stereo laparoscopic images using cone beam CT and fluoroscopy as bridging modalities. It does not require fiducials or manual alignment and compensates for tissue deformation from insufflation and respiration while allowing the laparoscope to be navigated. The paper’s theoretical and practical contributions are validated using simulated, phantom,

ex vivo

,

in vivo

and non medical data.

The Pico Lantern is proposed as a new tool for guidance in laparoscopic surgery. Its miniaturized design allows it to be picked up by a laparoscopic tool during surgery and tracked directly by the endoscope. By using laser projection, different patterns and annotations can be projected onto the tissue surface. The first explored application is surface reconstruction. The absolute error for surface reconstruction using stereo endoscopy and untracked Pico Lantern for a plane, cylinder and

ex vivo

kidney is 2.0 mm, 3.0 mm and 5.6 mm respectively. The absolute error using a mono endoscope and a tracked Pico Lantern for the same plane, cylinder and kidney is 0.8mm, 0.3mm and 1.5mm respectively. The results show the benefit of the wider baseline produced by tracking the Pico Lantern. Pulsatile motion of a human carotid artery is also detected

in vivo

. Future work will be done on the integration into standard and robot-assisted laparoscopic surgery.

Camera calibration is central to obtaining a quantitative image-to-physical-space mapping from stereo images acquired in the operating room (OR). A practical challenge for cameras mounted to the operating microscope is maintenance of image calibration as the surgeon’s field-of-view is repeatedly changed (in terms of zoom and focal settings) throughout a procedure. Here, we present an efficient method for sustaining a quantitative image-to-physical space relationship for arbitrary image acquisition settings (

S

) without the need for camera re-calibration. Essentially, we warp images acquired at

S

into the equivalent data acquired at a reference setting,

S

0

, using deformation fields obtained with optical flow by successively imaging a simple phantom. Closed-form expressions for the distortions were derived from which 3D surface reconstruction was performed based on the single calibration at

S

0

. The accuracy of the reconstructed surface was 1.05 mm and 0.59 mm along and perpendicular to the optical axis of the operating microscope on average, respectively, for six phantom image pairs, and was 1.26 mm and 0.71 mm for images acquired with a total of 47 arbitrary settings during three clinical cases. The technique is presented in the context of stereovision; however, it may also be applicable to other types of video image acquisitions (e.g., endoscope) because it does not rely on any

a priori

knowledge about the camera system itself, suggesting the method is likely of considerable significance.

A new concept of endoscopic device based on a compact optical probe which can capture 3D shape of objects using an active stereovision method is presented. The distinctive feature of this probe is its capability to dynamically switch between two distinct points of view. If the first measurement angle of view does not give results with sufficient quality, the system can switch to a second mode which sets distinct angle of view within less than 25 milliseconds. This feature consequently allows selecting the angle that provides the more useful 3D information and enhances the quality of the captured result.

The instrumental setup of this measurement system and the reconstruction algorithms are presented in this paper. Then, the advantages of this new endoscopic probe are explained with an experimental 3D reconstruction of a coin’s surface. Finally, first measurements on a phantom colon are provided. In future works, further miniaturization of the device and its integration into a real colonoscope will be implemented.

Many image-based systems for aiding the surgeon during minimally invasive surgery require the endoscopic camera to be calibrated at all times. This article proposes a method for accomplishing this goal whenever the camera has optical zoom and the focal length changes during the procedure. Our solution for online calibration builds on recent developments in tracking salient points using differential image alignment, is well suited for continuous operation, and makes no assumptions about the camera motion or scene rigidity. Experimental validation using both a phantom model and

in vivo

data shows that the method enables accurate estimation of focal length when the zoom varies, avoiding the need to explicitly recalibrate during surgery. To the best of our knowledge this the first work proposing a practical solution for online zoom calibration in the operation room.

Robust visual tracking of instruments is an important task in retinal microsurgery. In this context, the instruments are subject to a large variety of appearance changes due to illumination and other changes during a procedure, which makes the task very challenging. Most existing methods require collecting a sufficient amount of labelled data and yet perform poorly in handling appearance changes that are unseen in training data. To address these problems, we propose a new approach for robust instrument tracking. Specifically, we adopt an online learning technique that collects appearance samples of instruments on the fly and gradually learns a target-specific detector. Online learning enables the detector to reinforce its model and become more robust over time. The performance of the proposed method has been evaluated on a fully annotated dataset of retinal instruments in in-vivo retinal microsurgery and on a laparoscopy image sequence. In all experimental results, our proposed tracking approach shows superior performance compared to several other state-of-the-art approaches.

In this paper, we present the idea of equipping a tomographic medical scanner with a range imaging device (e.g. a 3D camera) to improve the current scanning workflow. A novel technical approach is proposed to robustly estimate patient surface geometry by a single snapshot from the camera. Leveraging the information of the patient surface geometry can provide significant clinical benefits, including automation of the scan, motion compensation for better image quality, sanity check of patient movement, augmented reality for guidance, patient specific dose optimization, and more. Our approach overcomes the technical difficulties resulting from suboptimal camera placement due to practical considerations. Experimental results on more than 30 patients from a real CT scanner demonstrate the robustness of our approach.

This paper presents a robotic control method for 3D steering of a beveled-tip flexible needle. The solution is based on a new duty-cycling control strategy that makes possible to control three degrees of freedom of the needle. A visual servoing control scheme using two orthogonal cameras observing a translucent phantom is then proposed to automatically steer a needle toward a 3D target point. Experimental results show a final positioning error of 0.4 mm and demonstrate the feasibility of this promising approach and its robustness to model errors.

Slipped Capital Femoral Epiphysis (SCFE) is a common hip displacement condition in adolescents. In the standard treatment, the surgeon uses intra-operative fluoroscopic imaging to plan the screw placement and the drill trajectory. The accuracy, duration, and efficacy of this procedure are highly dependent on surgeon skill. Longer procedure times result in higher radiation dose, to both patient and surgeon. A robotic system to guide the drill trajectory might help to reduce screw placement errors and procedure time by reducing the number of passes and confirmatory fluoroscopic images needed to verify accurate positioning of the drill guide along a planned trajectory. Therefore, with the long-term goals of improving screw placement accuracy, reducing procedure time and intra-operative radiation dose, our group is developing an image-guided robotic surgical system to assist a surgeon with pre-operative path planning and intra-operative drill guide placement.

Despite increased use of remote-controlled steerable catheter navigation systems for endovascular intervention, most current designs are based on master configurations which tend to alter natural operator tool interactions. This introduces problems to both ergonomics and shared human-robot control. This paper proposes a novel cooperative robotic catheterization system based on learning-from-demonstration. By encoding the higher-level structure of a catheterization task as a sequence of primitive motions, we demonstrate how to achieve prospective learning for complex tasks whilst incorporating subject-specific variations. A hierarchical Hidden Markov Model is used to model each movement primitive as well as their sequential relationship. This model is applied to generation of motion sequences, recognition of operator input, and prediction of future movements for the robot. The framework is validated by comparing catheter tip motions against the manual approach, showing significant improvements in the quality of catheterization. The results motivate the design of collaborative robotic systems that are intuitive to use, while reducing the cognitive workload of the operator.

The development of modern robotics and compact imaging detectors allows the transfer of diagnostic imaging modalities to the operating room, supporting surgeons to perform faster and safer procedures. An intervention that currently suffers from a lack of interventional imaging is radioembolization, a treatment for hepatic carcinoma. Currently, this procedure requires moving the patient from an angiography suite for preliminary catheterization and injection to a whole-body SPECT/CT for leakage detection, necessitating a second catheterization back in the angiography suite for the actual radioembolization. We propose an imaging setup that simplifies this procedure using a robotic approach to directly acquire an interventional SPECT/CT in the angiography suite. Using C-arm CT and a co-calibrated gamma camera mounted on a robotic arm, a personalized trajectory of the gamma camera is generated from the C-arm CT, enabling an interventional SPECT reconstruction that is inherently co-registered to the C-arm CT. In this work we demonstrate the feasibility of this personalized interventional SPECT/CT imaging approach in a liver phantom study.

Pectus excavatum is among the most common major congenital anomalies of the chest wall whose correction can be performed via minimally invasive Nuss technique that places a pectus bar to elevate the sternum anteriorly. However, the size and bending of the pectus bar are manually modeled intra-operatively by trial-and-error. The procedure requires intense pain management in the months following surgery. In response, we are developing a novel distraction device for incremental and personalized PE correction with minimal risk and pain, akin to orthodontic treatment using dental braces. To design the device, we propose in this study a personalized surgical planning framework for PE correction from clinical noncontrast CT. First, we segment the ribs and sternum via kernel graph cuts. Then costal cartilages, which have very low contrast in noncontrast CT, are modeled as 3D anatomical curves using the cosine series representation and estimated using a statistical shape model. The size and shape of the correction device are estimated through model fitting. Finally, the corrected/post-surgical chest is simulated in relation to the estimated shape of correction device. The root mean square mesh distance between the estimated cartilages and ground truth on 30 noncontrast CT scans was 1.28±0.81 mm. Our method found that the average deformation of the sterna and cartilages with the simulation of PE correction was 49.71±10.11 mm.

Automated Lymph Node (LN) detection is an important clinical diagnostic task but very challenging due to the low contrast of surrounding structures in Computed Tomography (CT) and to their varying sizes, poses, shapes and sparsely distributed locations. State-of-the-art studies show the performance range of 52.9% sensitivity at 3.1 false-positives per volume (FP/vol.), or 60.9% at 6.1 FP/vol. for mediastinal LN, by one-shot boosting on 3D HAAR features. In this paper, we first operate a preliminary candidate generation stage, towards ~100% sensitivity at the cost of high FP levels (~40 per patient), to harvest volumes of interest (VOI). Our 2.5D approach consequently decomposes any 3D VOI by resampling 2D reformatted orthogonal views

N

times, via scale, random translations, and rotations with respect to the VOI centroid coordinates. These random views are then used to train a deep Convolutional Neural Network (CNN) classifier. In testing, the CNN is employed to assign LN probabilities for all

N

random views that can be simply averaged (as a set) to compute the final classification probability per VOI. We validate the approach on two datasets: 90 CT volumes with 388 mediastinal LNs and 86 patients with 595 abdominal LNs. We achieve sensitivities of 70%/83% at 3 FP/vol. and 84%/90% at 6 FP/vol. in mediastinum and abdomen respectively, which drastically improves over the previous state-of-the-art work.

We propose a method to automatically select a treatment plan for radiotherapy of cervical cancer using a Plan-of-the-Day procedure, in which multiple treatment plans are constructed prior to treatment. The method comprises a multi-atlas based segmentation algorithm that uses the selected treatment plan to choose between two atlas sets. This segmentation only requires two registration procedures and can therefore be used in clinical practice without using excessive computation time. Our method is validated on a dataset of 224 treatment fractions for 10 patients. In 37 cases (16%), no recommendation was made by the algorithm due to poor image quality or registration results. In 93% of the remaining cases a correct recommendation for a treatment plan was given.

The radiographic appearance of breast tissue has been established as a strong risk factor for breast cancer. Here we present a complete machine learning framework for automatic estimation of mammographic density (MD) and robust feature extraction for breast cancer risk analysis. Our framework is able to simultaneously classify the breast region, fatty tissue, pectoral muscle, glandular tissue and nipple region. Integral to our method is the extraction of measures of breast density (as the fraction of the breast area occupied by glandular tissue) and mammographic pattern. A novel aspect of the segmentation framework is that a probability map associated with the label mask is provided, which indicates the level of confidence of each pixel being classified as the current label. The Pearson correlation coefficient between the estimated MD value and the ground truth is 0.8012 (p-value<0.0001). We demonstrate the capability of our methods to discriminate between women with and without cancer by analyzing the contralateral mammograms of 50 women with unilateral breast cancer, and 50 controls. Using MD we obtained an area under the ROC curve (AUC) of 0.61; however our texture-based measure of mammographic pattern significantly outperforms the MD discrimination with an AUC of 0.70.

Enlarged lymph nodes (LNs) can provide important information for cancer diagnosis, staging, and measuring treatment reactions, making automated detection a highly sought goal. In this paper, we propose a new algorithm representation of decomposing the LN detection problem into a set of 2D object detection subtasks on sampled CT slices, largely alleviating the curse of dimensionality issue. Our 2D detection can be effectively formulated as linear classification on a single image feature type of Histogram of Oriented Gradients (HOG), covering a moderate field-of-view of 45 by 45 voxels. We exploit both max-pooling and sparse linear fusion schemes to aggregate these 2D detection scores for the final 3D LN detection. In this manner, detection is more tractable and does not need to perform perfectly at instance level (as weak hypotheses) since our aggregation process will robustly harness collective information for LN detection. Two datasets (90 patients with 389 mediastinal LNs and 86 patients with 595 abdominal LNs) are used for validation. Cross-validation demonstrates 78.0% sensitivity at 6 false positives/volume (FP/vol.) (86.1% at 10 FP/vol.) and 73.1% sensitivity at 6 FP/vol. (87.2% at 10 FP/vol.), for the mediastinal and abdominal datasets respectively. Our results compare favorably to previous state-of-the-art methods.

Metastases to the lung are a therapeutic challenge because some are fast-evolving while others evolve slowly. Any insight that can be provided for which nodule has to be treated first would help clinicians. In this work, we evaluate the aggressiveness but also the response to treatment of these nodules using a calibrated mathematical model. This model is a macroscopic model describing tumoral growth through a set of nonlinear partial differential equations. It has to be calibrated to a specific patient and a specific nodule using a temporal sequence of CT scans. To this end, a new optimization technique based on a reduced order method is developed. Finally, results on two clinical cases are presented that give satisfactory numerical prognosis of the evolution of a nodule during different phases: growth, treatment and post-treatment relapse.

In this article, we describe a system for detecting dominant prostate tumors, based on a combination of features extracted from a novel multi-parametric quantitative ultrasound elastography technique. The performance of the system was validated on a data-set acquired from

n

 = 10 patients undergoing radical prostatectomy. Multi-frequency steady-state mechanical excitations were applied to each patient’s prostate through the perineum and prostate tissue displacements were captured by a transrectal ultrasound system. 3D volumetric data including absolute value of tissue elasticity, strain and frequency-response were computed for each patient. Based on the combination of all extracted features, a random forest classification algorithm was used to separate cancerous regions from normal tissue, and to compute a measure of cancer probability. Registered whole mount histopathology images of the excised prostate gland were used as a ground truth of cancer distribution for classifier training. An area under receiver operating characteristic curve of 0.82±0.01 was achieved in a leave-one-patient-out cross validation. Our results show the potential of multi-parametric quantitative elastography for prostate cancer detection for the first time in a clinical setting, and justify further studies to establish whether the approach can have clinical use.

Positron emission tomography computed tomography (PET-CT) is the preferred imaging modality for the evaluation of the lymphomas. Disease involvement in the lymphomas usually appear as foci of increased Fluorodeoxyglucose (FDG) uptake. Thresholding methods are applied to separate different regions of involvement. However, the main limitation of thresholding is that it also includes regions where there is normal FDG excretion and FDG uptake (NEUR) in structures such as the brain, bladder, heart and kidneys. We refer to these regions as NEURs (the normal excretion and uptake (of FDG) regions). NEURs can make image interpretation problematic. The ability to identify and label NEURs and separate them from abnormal regions is an important process that could improve the sensitivity of lesion detection and image interpretation. In this study, we propose a new method to automatically separate NEURs in thresholded PET images. We propose to group thresholded regions of the same structure with spatial and texture based clustering; we then classified NEURs on PET-CT contextual features. Our findings were that our approach had better accuracy when compared to conventional methods.

Until now, core needle biopsy of the axillary sentinel lymph nodes in early stage breast cancer patients is not possible, due to the lack of a proper combination of functional and anatomical information. In this work we present the first fully 3D freehand SPECT - ultrasound fusion, combining the advantages of both modalities. By using spatial positioning either with optical or with electromagnetic tracking for the ultrasound probe, and a mini gamma camera as radiation detector for freehand SPECT reconstructions, we investigate the capability of the introduced multi-model imaging system, where we compare both 3D freehand SPECT and 3D ultrasound to ground truth for a realistic breast mimicking phantom and further analyze the effect of tissue deformation by ultrasound. Finally, we also show its application in a real clinical setting.

Morphometric features characterizing the fusion and fragmentation of the glandular architecture of advanced prostate cancer have not previously been based upon the automated segmentation of discrete gland rings, due in part to the difficulty of extracting these structures from the H&E stained tissues. We present a novel approach for segmenting gland rings in multi-spectral immunofluorescence (IF) images and demonstrate the utility of the resultant features in predicting cancer recurrence in a cohort of 1956 images of prostate biopsies and prostatectomies from 679 patients. The proposed approach is evaluated for prediction of actual clinical outcomes of interest to physicians in comparison with previously published gland-unit features, yielding a concordance index (CI) of 0.67. This compares favorably to the CI of 0.66 obtained using a semi-automated segmentation of the corresponding H&E images from the same patients. This work presents the first algorithms for segmentation of gland rings lacking a central lumen, and for separation of touching epithelial units, and introduces new gland adjacency features for predicting prostate cancer clinical progression across both biopsy and prostatectomy images.

Detection of new or rapidly evolving melanocytic lesions is crucial for early diagnosis and treatment of melanoma. We propose a fully automated pre-screening system for detecting new lesions or changes in existing ones, on the order of 2 − 3mm, over almost the entire body surface. Our solution is based on a multi-camera 3D stereo system. The system captures 3D textured scans of a subject at different times and then brings these scans into correspondence by aligning them with a learned, parametric, non-rigid 3D body model. This means that captured skin textures are in accurate alignment across scans, facilitating the detection of new or changing lesions. The integration of lesion segmentation with a deformable 3D body model is a key contribution that makes our approach robust to changes in illumination and subject pose.

Bone tumor segmentation on bone scans has recently been adopted as a basis for objective tumor assessment in several phase II and III clinical drug trials. Interpretation can be difficult due to the highly sensitive but non-specific nature of bone tumor appearance on bone scans. In this paper we present a machine learning approach to segmenting tumors on bone scans, using intensity and context features aimed at addressing areas prone to false positives. We computed the context features using landmark points, identified by a modified active shape model. We trained a random forest classifier on 100 and evaluated on 73 prostate cancer subjects from a multi-center clinical trial. A reference segmentation was provided by a board certified radiologist. We evaluated our learning based method using the Jaccard index and compared against the state of the art, rule based method. Results showed an improvement from 0.50 ±0.31 to 0.57 ±0.27. We found that the context features played a significant role in the random forest classifier, helping to correctly classify regions prone to false positives.

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is a powerful protocol for assessing tumour progression from changes in tissue contrast enhancement. Manual colorectal tumour delineation is a challenging and time consuming task due to the complex enhancement patterns in the 4D sequence. There is a need for a consistent approach to colorectal tumour segmentation in DCE-MRI and we propose a novel method based on detection of the tumour from signal enhancement characteristics of homogeneous tumour subregions and their neighbourhoods. Our method successfully detected 20 of 23 cases with a mean Dice score of 0.68 ±0.15 compared to expert annotations, which is not significantly different from expert inter-rater variability of 0.73 ±0.13 and 0.77 ±0.10. In comparison, a standard DCE-MRI tumour segmentation technique, fuzzy c-means, obtained a Dice score of 0.28 ±0.17.

Internal examinations such as Digital Rectal Examination (DRE) and bimanual Vaginal Examination (BVE) are routinely performed for early diagnosis of cancer and other diseases. Although they are recognised as core skills to be taught on a medical curriculum, they are difficult to learn and teach due to their unsighted nature. We present a framework that combines a visualisation and analysis tool with position and pressure sensors to enable the study of internal examinations and provision of real-time feedback. This approach is novel as it allows for real-time continuous trajectory and pressure data to be obtained for the complete examination, which may be used for teaching and assessment. Experiments were conducted performing DRE and BVE on benchtop models, and BVE on Gynaecological Teaching Assistants (GTA). The results obtained suggest that the proposed methodology may provide an insight into what constitutes an adequate DRE or BVE, provide real-time feedback tools for learning and assessment, and inform haptics-based simulator design.

This paper aims to trace retinal blood vessel trees in fundus images. This task is far from being trivial as the

crossover

of vessels are commonly encountered in image-based vessel networks. Meanwhile it is often crucial to separate the vessel tree structures in applications such as diabetic retinopathy analysis. In this work, a novel directed graph based approach is proposed to cast the task as label propagation over directed graphs, such that the graph is to be partitioned into disjoint sub-graphs, or equivalently, each of the vessel trees is traced and separated from the rest of the vessel network. Then the tracing problem is addressed by making novel usage of the matrix-forest theorem in algebraic graph theory. Empirical experiments on synthetic as well as publicly available fundus image datasets demonstrate the applicability of our approach.

In this work, we present a novel method for blood vessel segmentation in fundus images based on a discriminatively trained, fully connected conditional random field model. Retinal image analysis is greatly aided by blood vessel segmentation as the vessel structure may be considered both a key source of signal, e.g. in the diagnosis of diabetic retinopathy, or a nuisance, e.g. in the analysis of pigment epithelium or choroid related abnormalities. Blood vessel segmentation in fundus images has been considered extensively in the literature, but remains a challenge largely due to the desired structures being thin and elongated, a setting that performs particularly poorly using standard segmentation priors such as a Potts model or total variation. In this work, we overcome this difficulty using a discriminatively trained conditional random field model with more expressive potentials. In particular, we employ recent results enabling extremely fast inference in a fully connected model. We find that this rich but computationally efficient model family, combined with principled discriminative training based on a structured output support vector machine yields a fully automated system that achieves results statistically indistinguishable from an expert human annotator. Implementation details are available at http://pages.saclay.inria.fr/ matthew.blaschko/projects/retina/.

Optical colonoscopy is the preferred modality for the screening and prevention of colorectal cancer. Chromoendoscopy can increase lesion detection rate by highlighting tissue topography with a colored dye, but is too time-consuming to be adopted in routine colonoscopy screening. We developed a fast and dye-free technique that generates virtual chromoendoscopy images that incorporate topography features acquired from photometric stereo endoscopy. We demonstrate that virtual chromoendoscopy augmented by topography achieves similar image quality to conventional chromoendoscopy in ex-vivo swine colon.

This paper proposes a novel super-resolution framework to reconstruct high-resolution fundus images from multiple low-resolution video frames in retinal fundus imaging. Natural eye movements during an examination are used as a cue for super-resolution in a robust maximum a-posteriori scheme. In order to compensate heterogeneous illumination on the fundus, we integrate retrospective illumination correction for photometric registration to the underlying imaging model. Our method utilizes quality self-assessment to provide objective quality scores for reconstructed images as well as to select regularization parameters automatically. In our evaluation on real data acquired from six human subjects with a low-cost video camera, the proposed method achieved considerable enhancements of low-resolution frames and improved noise and sharpness characteristics by 74%. In terms of image analysis, we demonstrate the importance of our method for the improvement of automatic blood vessel segmentation as an example application, where the sensitivity was increased by 13% using super-resolution reconstruction.

Nailfold capillaroscopy is an established qualitative technique in the assessment of patients displaying Raynaud’s phenomenon. We describe a fully automated system for extracting quantitative biomarkers from capillaroscopy images, using a layered machine learning approach. On an unseen set of 455 images, the system detects and locates individual capillaries as well as human experts, and makes measurements of vessel morphology that reveal statistically significant differences between patients with (relatively benign) primary Raynaud’s phenomenon, and those with potentially life-threatening systemic sclerosis.

Label propagation has been shown to be effective in many automatic segmentation applications. However, its reliance on accurate image alignment means that segmentation results can be affected by any registration errors which occur. Patch-based methods relax this dependence by avoiding explicit one-to-one correspondence assumptions between images but are still limited by the search window size. Too small, and it does not account for enough registration error; too big, and it becomes more likely to select incorrect patches of similar appearance for label fusion. This paper presents a novel patch-based label propagation approach which uses relative geodesic distances to define patient-specific coordinate systems as spatial context to overcome this problem. The approach is evaluated on multi-organ segmentation of 20 cardiac MR images and 100 abdominal CT images, demonstrating competitive results.

Model-based approaches are very popular for medical image segmentation as they carry useful prior information on the target structure. Among them, the implicit template deformation framework recently bridged the gap between the efficiency and flexibility of level-set region competition and the robustness of atlas deformation approaches. This paper generalizes this method by introducing the notion of tagged templates. A tagged template is an implicit model in which different subregions are defined. In each of these subregions, specific image features can be used with various confidence levels. The tags can be either set manually or automatically learnt via a process also hereby described. This generalization therefore greatly widens the scope of potential clinical application of implicit template deformation while maintaining its appealing algorithmic efficiency. We show the great potential of our approach in myocardium segmentation of ultrasound images.

Accurate automated segmentation of the right ventricle is difficult due in part to the large shape variation found between patients. We explore the ability of manifold learning based shape models to represent the complexity of shape variation found within an RV dataset as compared to a typical PCA based model. This is empirically evaluated with the manifold model displaying a greater ability to represent complex shapes. Furthermore, we present a combined manifold shape model and Markov Random Field Segmentation framework. The novelty of this method is the iterative generation of targeted shape priors from the manifold using image information and a current estimate of the segmentation; a process that can be seen as a traversal across the manifold. We apply our method to the independently evaluated MICCAI 2012 RV Segmentation Challenge data set. Our method performs similarly or better than the state-of-the-art methods.

Graph-cut algorithms have been extensively investigated for interactive binary segmentation, when the simultaneous delineation of multiple objects can save considerable user’s time. We present an algorithm (named DRIFT) for 3D multiple object segmentation based on seed voxels and Differential Image Foresting Transforms (DIFTs) with relaxation. DRIFT stands behind efficient implementations of some state-of-the-art methods. The user can add/remove markers (seed voxels) along a sequence of executions of the DRIFT algorithm to improve segmentation. Its first execution takes linear time with the image’s size, while the subsequent executions for corrections take sublinear time in practice. At each execution, DRIFT first runs the DIFT algorithm, then it applies diffusion filtering to smooth boundaries between objects (and background) and, finally, it corrects possible objects’ disconnection occurrences with respect to their seeds. We evaluate DRIFT in 3D CT-images of the thorax for segmenting the arterial system, esophagus, left pleural cavity, right pleural cavity, trachea and bronchi, and the venous system.

Delineation and noise removal play a significant role in clinical quantification of PET images. Conventionally, these two tasks are considered independent, however, denoising can improve the performance of boundary delineation by enhancing SNR while preserving the structural continuity of local regions. On the other hand, we postulate that segmentation can help denoising process by constraining the smoothing criteria locally. Herein, we present a novel iterative approach for simultaneous PET image denoising and segmentation. The proposed algorithm uses generalized Anscombe transformation priori to non-local means based noise removal scheme and affinity propagation based delineation. For non-local means denoising, we propose a new regional means approach where we automatically and efficiently extract the appropriate subset of the image voxels by incorporating the class information from affinity propagation based segmentation. PET images after denoising are further utilized for refinement of the segmentation in an iterative manner. Qualitative and quantitative results demonstrate that the proposed framework successfully removes the noise from PET images while preserving the structures, and improves the segmentation accuracy.

Magnetic resonance guided high intensity focused ultrasound (MRgHIFU) is a new type of minimally invasive therapy for treating malignant liver tissues. Since the ribs on the beam path can compromise an effective therapy, detecting them and tracking their motion on MR images is of great importance. However, due to poor magnetic signal emission of bones, ribs cannot be entirely observed in MR. In the proposed method, we take advantage of the accuracy of CT in imaging the ribs to build a geometric ribcage model and combine it with an appearance model of the neighbouring structures of ribs in MR to reconstruct realistic centerlines in MRIs. We have improved our previous method by using a more sophisticated appearance model, a more flexible ribcage model, and a more effective optimization strategy. We decreased the mean error to 2.5 mm, making the method suitable for clinical application. Finally, we propose a rib registration method which conserves the shape and length of ribs, and imposes realistic constraints on their motions, achieving 2.7 mm mean accuracy.

In contrast to preoperative brain tumor segmentation, the problem of postoperative brain tumor segmentation has been rarely approached so far. We present a fully-automatic segmentation method using multimodal magnetic resonance image data and patient-specific semi-supervised learning. The idea behind our semi-supervised approach is to effectively fuse information from both pre- and postoperative image data of the same patient to improve segmentation of the postoperative image. We pose image segmentation as a classification problem and solve it by adopting a semi-supervised decision forest. The method is evaluated on a cohort of 10 high-grade glioma patients, with segmentation performance and computation time comparable or superior to a state-of-the-art brain tumor segmentation method. Moreover, our results confirm that the inclusion of preoperative MR images lead to a better performance regarding postoperative brain tumor segmentation.

Cortical thickness (CT) is an important morphometric measure that has implications for psychiatric and neurologic processes. We propose a novel approach for automatically computing CT in an accurate and robust manner using LOGISMOS-B: Layered Optimal Graph Image Segmentation of Multiple Objects and Surfaces for the Brain. LOGISMOS-B is a cortical surface segmentation method based on LOGISMOS graph segmentation and generalized gradient vector flows. We evaluate our method on two different datasets (

n

 = 83 total). The results show that LOGISMOS-B is more accurate than the popular FreeSurfer (FS) method and provides more reliable thickness measurements across a variety of challenging images. LOGISMOS-B accurately recovers known CT patterns, both across cortical lobes and locally, such as between the banks of the central sulcus, in healthy subjects and MS patients. Manual landmarks indicate a signed surface distance of 0.081±0.447mm for WM and 0.018±0.498mm for LOGISMOS-B, compared to 0.263±0.452mm for WM and − 0.167±0.556mm for GM for FS, highlighting the surface placement accuracy of LOGISMOS-B. Finally, a regresion study shows that LOGISMOS-B provides strong correlation with age and plausible annual thinning rates across the cortex, with locally discerning thinning patterns, in agreement with the literature.

In this paper, we present a novel label inference method that integrates registration and patch priors, and serves as a remedy for labelling errors around structural boundaries. With the initial label map provided by nonrigid registration methods, its corresponding signed distance function can be estimated and used to evaluate the segmentation confidence. The pixels with less confident labels are selected as candidate nodes to be refined and those with relatively confident results are settled as seeds. The affinity between seeds and candidate nodes, which consists of regular image lattice connections, registration prior based on signed distance and patch prior from the warped atlas, is encoded to guide the label inference procedure. For method evaluation, experiments have been carried out on two publicly available data sets and it only takes several seconds for our method to improve the segmentation quality significantly.

The need to segment multiple interacting surfaces is a common problem in medical imaging and it is often assumed that such surfaces are continuous within the confines of the region of interest. However, in some application areas, the surfaces of interest may contain a shared hole in which the surfaces no longer exist and the exact location of the hole boundary is not known

a priori

. The boundary of the neural canal opening seen in spectral-domain optical coherence tomography volumes is an example of a “hole” embedded with multiple surrounding surfaces. Segmentation approaches that rely on finding the surfaces alone are prone to failures as deeper structures within the hole can “attract” the surfaces and pull them away from their correct location at the hole boundary. With this application area in mind, we present a graph-theoretic approach for segmenting multiple surfaces with a shared hole. The overall cost function that is optimized consists of both the costs of the surfaces outside the hole and the cost of boundary of the hole itself. The constraints utilized were appropriately adapted in order to ensure the smoothness of the hole boundary in addition to ensuring the smoothness of the non-overlapping surfaces. By using this approach, a significant improvement was observed over a more traditional two-pass approach in which the surfaces are segmented first (assuming the presence of no hole) followed by segmenting the neural canal opening.

We present a novel framework for

depth

based optic cup boundary extraction from a

single

2D color fundus photograph per eye. Multiple depth estimates from shading, color and texture gradients in the image are correlated with Optical Coherence Tomography (OCT) based depth using a coupled sparse dictionary, trained on image-depth pairs. Finally, a Markov Random Field is formulated on the depth map to model the relative depth and discontinuity at the cup boundary. Leave-one-out validation of depth estimation on the

INSPIRE

dataset gave average correlation coefficient of 0.80. Our cup segmentation outperforms several state-of-the-art methods on the

DRISHTI-GS

dataset with an average F-score of 0.81 and boundary-error of 21.21 pixels on test set against manual expert markings. Evaluation on an additional set of 28 images against OCT scanner provided groundtruth showed an average rms error of 0.11 on Cup-Disk diameter and 0.19 on Cup-disk area ratios.

One of the main problems of the existing methods for the segmentation of cerebral vasculature is the appearance in the segmentation result of wrong topological artefacts such as the kissing vessels. In this paper, a new approach for the detection and correction of such errors is presented. The proposed technique combines robust topological information given by Persistent Homology with complementary geometrical information of the vascular tree. The method was evaluated on 20 images depicting cerebral arteries. Detection and correction success rates were 81.80% and 68.77%, respectively.

In this paper, we propose a new method for simultaneously segmenting brain scans of glioma patients and registering these scans to a normal atlas. Performing joint segmentation and registration for brain tumors is very challenging when tumors include multifocal masses and have complex shapes with heterogeneous textures. Our approach grows tumors for each mass from multiple seed points using a tumor growth model and modifies a normal atlas into one with tumors and edema using the combined results of grown tumors. We also generate a tumor shape prior via the random walk with restart, utilizing multiple tumor seeds as initial foreground information. We then incorporate this shape prior into an EM framework which estimates the mapping between the modified atlas and the scans, posteriors for each tissue labels, and the tumor growth model parameters. We apply our method to the BRATS 2013 leaderboard dataset to evaluate segmentation performance. Our method shows the best performance among all participants.

We propose a fast algorithm to estimate brain tissue concentrations from conventional T1-weighted images based on a Bayesian maximum a posteriori formulation that extends the “mixel” model developed in the 90’s. A key observation is the necessity to incorporate additional prior constraints to the “mixel” model for the estimation of plausible concentration maps. Experiments on the ADNI standardized dataset show that global and local brain atrophy measures from the proposed algorithm yield enhanced diagnosis testing value than with several widely used soft tissue labeling methods.

Interventional cardiologists are often challenged by a high degree of variability in the coronary venous anatomy during coronary sinus cannulation and left ventricular epicardial lead placement for cardiac resynchronization therapy (CRT), making it important to have a precise and fully-automatic segmentation solution for detecting the coronary sinus. A few approaches have been proposed for automatic segmentation of tubular structures utilizing various vesselness measurements. Although working well on contrasted coronary arteries, these methods fail in segmenting the coronary sinus that has almost no contrast in computed tomography angiography (CTA) data, making it difficult to distinguish from surrounding tissues. In this work we propose a multiscale sparse appearance learning based method for estimating vesselness towards automatically extracting the centerlines. Instead of modeling the subtle discrimination at the low-level intensity, we leverage the flexibility of sparse representation to model the inherent spatial coherence of vessel/background appearance and derive a vesselness measurement. After centerline extraction, the coronary sinus lumen is segmented using a learning based boundary detector and Markov random field (MRF) based optimal surface extraction. Quantitative evaluation on a large cardiac CTA dataset (consisting of 204 3D volumes) demonstrates the superior accuracy of the proposed method in both centerline extraction and lumen segmentation, compared to the state-of-the-art.

We present an unsupervised approach to segment optic cups in fundus images for glaucoma detection without using any additional training images. Our approach follows the superpixel framework and domain prior recently proposed in [1], where the superpixel classification task is formulated as a low-rank representation (LRR) problem with an efficient closed-form solution. Moreover, we also develop an adaptive strategy for automatically choosing the only parameter in LRR and obtaining the final result for each image. Evaluated on the popular

ORIGA

dataset, the results show that our approach achieves better performance compared with existing techniques.

An efficient and accurate segmentation of 3D transrectal ultrasound (TRUS) images plays an important role in the planning and treatment of the practical 3D TRUS guided prostate biopsy. However, a meaningful segmentation of 3D TRUS images tends to suffer from US speckles, shadowing and missing edges etc, which make it a challenging task to delineate the correct prostate boundaries. In this paper, we propose a novel convex optimization based approach to extracting the prostate surface from the given 3D TRUS image, while preserving a new global volume-size prior. We, especially, study the proposed combinatorial optimization problem by convex relaxation and introduce its dual continuous max-flow formulation with the new bounded flow conservation constraint, which results in an efficient numerical solver implemented on GPUs. Experimental results using 12 patient 3D TRUS images show that the proposed approach while preserving the volume-size prior yielded a mean DSC of 89.5%±2.4%, a MAD of 1.4±0.6

mm

, a MAXD of 5.2±3.2

mm

, and a VD of 7.5%±6.2% in ~1 minute, deomonstrating the advantages of both accuracy and efficiency. In addition, the low standard deviation of the segmentation accuracy shows a good reliability of the proposed approach.

The diversity in appearance of diseased lung tissue makes automatic segmentation of lungs from CT with severe pathologies challenging. To overcome this challenge, we rely on contextual constraints from neighboring anatomies to detect and segment lung tissue across a variety of pathologies. We propose an algorithm that combines statistical learning with these anatomical constraints to seek a segmentation of the lung consistent with adjacent structures, such as the heart, liver, spleen, and ribs. We demonstrate that our algorithm reduces the number of failed detections and increases the accuracy of the segmentation on unseen test cases with severe pathologies.

Errata List

Page

Corrections

180

Algorithm 1, Line 5: The denominator term in the expression for estimating

σ

2

should be 3

n

rather than

n

, i.e.,

σ

2

is the average square residual distance of the match positions divided by the spatial dimensionality:

$\sigma^2 = \frac{1}{3n}\displaystyle\sum^n_{i=1}\| {y_{\text{p}i}}- T({x_{\text{p}i}}) \|_2^2$

184

Figure 2: Please see below for the corrected version of this figure. A bug in the script used to generate the original figure resulted in mis-plotting the values corresponding to the IMLOP algorithm. Note that the corrected figure strengthens the conclusions of the paper, i.e., the registration accuracy of IMLOP is better than previously shown and IMLOP detects the inaccurate registration outcomes more robustly than previously shown.

Fig. 2.

Corrected version of Figure 2