Medical Image Computing and Computer-Assisted Intervention – MICCAI 2010

MR image data can provide many features or measures although any single measure is unlikely to comprehensively characterize the underlying morphology. We present a framework in which multiple measures are used in manifold learning steps to generate coordinate embeddings which are then combined to give an improved single representation of the population. An application to neonatal brain MRI data shows that the use of shape and appearance measures in particular leads to biologically plausible and consistent representations correlating well with clinical data. Orthogonality among the correlations suggests the embedding components relate to comparatively independent morphological features. The rapid changes that occur in brain shape and in MR image appearance during neonatal brain development justify the use of shape measures (obtained from a deformation metric) and appearance measures (obtained from image similarity). The benefit of combining separate embeddings is demonstrated by improved correlations with clinical data and we illustrate the potential of the proposed framework in characterizing trajectories of brain development.

Updating segmentation results in real-time based on repeated user input is a reliable way to guarantee accuracy, paramount in medical imaging applications, while making efficient use of an expert’s time. The random walker algorithm with priors is a robust method able to find a globally optimal probabilistic segmentation with an intuitive method for user input. However, like many other segmentation algorithms, it can be too slow for real-time user interaction. We propose a speedup to this popular algorithm based on offline precomputation, taking advantage of the time images are stored on servers prior to an analysis session. Our results demonstrate the benefits of our approach. For example, the segmentations found by the original random walker and by our new precomputation method for a given 3D image have a Dice’s similarity coefficient of 0.975, yet our method runs in 1/25

th

of the time.

Changes in corpus callosum (CC) size are typically quantified in clinical studies by measuring the CC cross-sectional area on a midsagittal plane. We propose an alternative measurement plane based on the role of the CC as a bottleneck structure in determining the rate of interhemispheric neural transmission. We designate this plane as the Minimum Corpus Callosum Area Plane (MCCAP), which captures the cross section of the CC that best represents an upper bound on interhemispheric transmission. Our MCCAP extraction method uses a nested optimization framework, segmenting the CC as it appears on each candidate plane, using registration-based segmentation. We demonstrate the robust convergence and high accuracy of our method for magnetic resonance images and present preliminary clinical results showing higher sensitivity to disease-induced atrophy.

In order to evaluate the quality of segmentations of an image and assess intra- and inter-expert variability in segmentation performance, an Expectation Maximization (EM) algorithm for Simultaneous Truth And Performance Level Estimation (STAPLE) was recently developed. This algorithm, originally presented for segmentation validation, has since been used for many applications, such as atlas construction and decision fusion. However, the manual delineation of structures of interest is a very time consuming and burdensome task. Further, as the time required and burden of manual delineation increase, the accuracy of the delineation is decreased. Therefore, it may be desirable to ask the experts to delineate only a reduced number of structures or the segmentation of all structures by all experts may simply not be achieved. Fusion from data with some structures not segmented by each expert should be carried out in a manner that accounts for the missing information. In other applications, locally inconsistent segmentations may drive the STAPLE algorithm into an undesirable local optimum, leading to misclassifications or misleading experts performance parameters.

We present a new algorithm that allows fusion with partial delineation and which can avoid convergence to undesirable local optima in the presence of strongly inconsistent segmentations. The algorithm extends STAPLE by incorporating prior probabilities for the expert performance parameters. This is achieved through a Maximum A Posteriori formulation, where the prior probabilities for the performance parameters are modeled by a beta distribution. We demonstrate that this new algorithm enables dramatically improved fusion from data with partial delineation by each expert in comparison to fusion with STAPLE.

We present a method for automatically segmenting the blood vessels in optic nerve head (ONH) centered spectral-domain optical coherence tomography (SD-OCT) volumes, with a focus on the ability to segment the vessels in the region near the neural canal opening (NCO). The algorithm first pre-segments the NCO using a graph-theoretic approach. Oriented Gabor wavelets rotated around the center of the NCO are applied to extract features in a 2-D vessel-aimed projection image. Corresponding oriented NCO-based templates are utilized to help suppress the false positive tendency near the NCO boundary. The vessels are identified in a vessel-aimed projection image using a pixel classification algorithm. Based on the 2-D vessel profiles, 3-D vessel segmentation is performed by a triangular-mesh-based graph search approach in the SD-OCT volume. The segmentation method is trained on 5 and is tested on 10 randomly chosen independent ONH-centered SD-OCT volumes from 15 subjects with glaucoma. Using ROC analysis, for the 2-D vessel segmentation, we demonstrate an improvement over the closest previous work with an area under the curve (AUC) of 0.81 (0.72 for previously reported approach) for the region around the NCO and 0.84 for the region outside the NCO (0.81 for previously reported approach).

Identification of Gad-enhancing lesions is of great interest in Multiple Sclerosis (MS) disease since they are associated with disease activity. Current techniques for detecting Gad-enhancing lesions use a contrast agent (Gadolinium) which is administered intravenously to highlight Gad-enhancing lesions. However, the contrast agent not only highlights these lesions, but also causes other tissues (e.g. blood vessels) or noise in the Magnetic Resonance Image (MRI) to appear hyperintense. Discrimination of enhanced lesions from other enhanced structures is particularly challenging as these lesions are typically small and can be found in close proximity to vessels. We present a new approach to address the segmentation of Gad-enhancing MS lesions using Conditional Random Fields (CRF). CRF performs the classification by simultaneously incorporating the spatial dependencies of data and labels. The performance of the CRF classifier on 20 clinical data sets shows promising results in successfully capturing all Gad-enhancing lesions. Furthermore, the quantitative results of the CRF classifier indicate a reduction in the False Positive (FP) rate by an average factor of 5.8 when comparing to Linear Discriminant Analysis (LDA) and 1.6 comparing to a Markov Random Field (MRF) classifier.

In this paper we propose a novel and robust system for the automated identification of major sulci on cortical surfaces. Using multiscale representation and intrinsic surface mapping, our system encodes anatomical priors in manually traced sulcal lines with an intrinsic atlas of major sulci. This allows the computation of both individual and joint likelihood of sulcal lines for their automatic identification on cortical surfaces. By modeling sulcal anatomy with intrinsic geometry, our system is invariant to pose differences and robust across populations and surface extraction methods. In our experiments, we present quantitative validations on twelve major sulci to show the excellent agreement of our results with manually traced curves. We also demonstrate the robustness of our system by successfully applying an atlas of Chinese population to identify sulci on Caucasian brains of different age groups, and surfaces extracted by three popular software tools.

We present a new approach for quantifying the degradation of knee cartilage in the medial meniscal tear (MMT) model of osteoarthritis in the rat. A statistical strategy was used to guide the selection of a region of interest (ROI) from the images obtained from a pilot study. We hypothesize that this strategy can be used to localize a region of cartilage most vulnerable to MMT-induced damage. In order to test this hypothesis, a longitudinal study was conducted in which knee cartilage thickness in a pre-selected ROI was monitored for three weeks and comparisons were made between MMT and control rats. We observed a significant decrease in cartilage thickness in MMT rats and a significant increase in cartilage thickness in sham-operated rats as early as one week post surgery when compared to pre-surgery measurements.

This paper proposes a novel technique for constructing a neuroanatomical

shape complex

atlas using an information geometry framework. A shape complex is a collection of shapes in a local neighborhood. We represent the boundary of the entire shape complex using the zero level set of a distance function

$S(\textbf{x})$

. The spatial relations between the different anatomical structures constituting the shape complex are captured via the distance transform. We then leverage the well known relationship between the stationary state wave function

ψ

(

x

) of the Schrödinger equation

$-\hbar^2\nabla^2 \psi+\psi=0$

and the eikonal equation

$\|\nabla S\|=1$

satisfied by any distance function

$S(\textbf{x})$

. This leads to a one-to-one map between

$\psi(\textbf{x})$

and

S

(

x

) and allows for recovery of

$S(\textbf{x})$

from

ψ

(

x

) through an explicit mathematical relationship. Since the wave function can be regarded as a square-root density function, we are able to exploit this connection and convert shape complex distance transforms into probability density functions. Furthermore, square-root density functions can be seen as points on a unit hypersphere whose Riemannian structure is fully known. A shape complex atlas is constructed by first computing the Karcher mean

$\bar\psi(\textbf{x})$

of the wave functions, followed by an inverse mapping of the estimated mean back to the space of distance transforms in order to realize the atlas. We demonstrate the shape complex atlas computation via a set of experiments on a population of brain MRI scans. We also present modes of variation from the computed atlas for the control population to demonstrate the shape complex variability.

We present a new non-parametric model constraint graph min-cut algorithm for automatic kidney segmentation in CT images. The segmentation is formulated as a maximum a-posteriori estimation of a model-driven Markov random field. A non-parametric hybrid shape and intensity model is treated as a latent variable in the energy functional. The latent model and labeling map that minimize the energy functional are then simultaneously computed with an expectation maximization approach. The main advantages of our method are that it does not assume a fixed parametric prior model, which is subjective to inter-patient variability and registration errors, and that it combines both the model and the image information into a unified graph min-cut based segmentation framework. We evaluated our method on 20 kidneys from 10 CT datasets with and without contrast agent for which ground-truth segmentations were generated by averaging three manual segmentations. Our method yields an average volumetric overlap error of 10.95%, and average symmetric surface distance of 0.79mm. These results indicate that our method is accurate and robust for kidney segmentation.

Additional material can be found at

http://www.cs.huji.ac.il/~freiman/kidney_seg

.

From the image analysis perspective, a disadvantage of MRI is the lack of image intensity standardization. Differences in coil sensitivity, pulse sequence and acquisition parameters lead to very different mappings from tissue properties to image intensity levels. This presents challenges for image analysis techniques because the distribution of image intensities for different brain regions can change substantially from scan to scan. Though intensity correction can sometimes alleviate this problem, it fails in more difficult scenarios in which different types of tissue are mapped to similar gray levels in one scan but different intensities in another. Here, we propose using multi-spectral data to create synthetic MRI scans matched to the intensity distribution of a given dataset using a physical model of acquisition. If the multi-spectral data are manually annotated, the labels can be transfered to the synthetic scans to build a dataset-tailored gold standard. The approach was tested on a multi-atlas based hippocampus segmentation framework using a publicly available database, significantly improving the results obtained with other intensity correction methods.

The interpretation of medical images benefits from anatomical and physiological priors to optimize computer-aided diagnosis (CAD) applications. Diagnosis also relies on the comprehensive analysis of multiple organs and quantitative measures of soft tissue. An automated method optimized for medical image data is presented for the simultaneous segmentation of four abdominal organs from 4D CT data using graph cuts. Contrast-enhanced CT scans were obtained at two phases: non-contrast and portal venous. Intra-patient data were spatially normalized by non-linear registration. Then 4D erosion using population historic information of contrast-enhanced liver, spleen, and kidneys was applied to multi-phase data to initialize the 4D graph and adapt to patient specific data. CT enhancement information and constraints on shape, from Parzen windows, and location, from a probabilistic atlas, were input into a new formulation of a 4D graph. Comparative results demonstrate the effects of appearance and enhancement, and shape and location on organ segmentation.

Quantitative information about the geometry of the carotid artery bifurcation may help in predicting the development of atherosclerosis. A geodesic active contours based segmentation method combining both gradient and intensity information was developed for semi-automatic, accurate and robust quantification of the carotid bifurcation angle in Black Blood MRA data. The segmentation method was evaluated by comparing its accuracy to inter and intra observer variability on a large dataset that has been acquired as part of a longitudinal population study which investigates the natural progression of carotid atherosclerosis. Furthermore, the method is shown to be robust to initialization differences. The bifurcation angle obtained from the segmented lumen corresponds well with the angle derived from the manual lumen segmentation, which demonstrates that the method has large potential to replace manual segmentations for extracting the carotid bifurcation angle from Black Blood MRA data.

We propose a simple strategy to improve automatic medical image segmentation. The key idea is that without deep understanding of a segmentation method, we can still improve its performance by directly calibrating its results with respect to manual segmentation. We formulate the calibration process as a bias correction problem, which is addressed by machine learning using training data. We apply this methodology on three segmentation problems/methods and show significant improvements for all of them.

We address the problem of 3-D Mesh segmentation for categories of objects with known part structure. Part labels are derived from a semantic interpretation of non-overlapping subsurfaces. Our approach models the label distribution using a Conditional Random Field (CRF) that imposes constraints on the relative spatial arrangement of neighboring labels, thereby ensuring semantic consistency. To this end, each label variable is associated with a rich shape descriptor that is intrinsic to the surface. Randomized decision trees and cross validation are employed for learning the model, which is eventually applied using graph cuts. The method is flexible enough for segmenting even geometrically less structured regions and is robust to local and global shape variations.

Clinical trials of anti-angiogenic and vascular-disrupting agents often use biomarkers derived from DCE-MRI, typically reporting whole-tumor summary statistics and so overlooking spatial parameter variations caused by tissue heterogeneity. We present a data-driven segmentation method comprising tracer-kinetic model-driven registration for motion correction, conversion from MR signal intensity to contrast agent concentration for cross-visit normalization, iterative principal components analysis for imputation of missing data and dimensionality reduction, and statistical outlier detection using the minimum covariance determinant to obtain a robust Mahalanobis distance. After applying these techniques we cluster in the principal components space using

k

-means. We present results from a clinical trial of a VEGF inhibitor, using time-series data selected because of problems due to motion and outlier time series. We obtained spatially-contiguous clusters that map to regions with distinct microvascular characteristics. This methodology has the potential to uncover localized effects in trials using DCE-MRI-based biomarkers.

Quantitative magnetic resonance analysis often requires accurate, robust and reliable automatic extraction of anatomical structures. Recently, template-warping methods incorporating a label fusion strategy have demonstrated high accuracy in segmenting cerebral structures. In this study, we propose a novel patch-based method using expert segmentation priors to achieve this task. Inspired by recent work in image denoising, the proposed nonlocal patch-based label fusion produces accurate and robust segmentation. During our experiments, the hippocampi of 80 healthy subjects were segmented. The influence on segmentation accuracy of different parameters such as patch size or number of training subjects was also studied. Moreover, a comparison with an appearance-based method and a template-based method was carried out. The highest median kappa value obtained with the proposed method was 0.884, which is competitive compared with recently published methods.

In this paper, we re-examine the cellular automata(CA) algorithm to show that the result of its state evolution converges to that of the shortest path algorithm. We proposed a complete tumor segmentation method on post contrast T1 MR images, which standardizes the VOI and seed selection, uses CA transition rules adapted to the problem and evolves a level set surface on CA states to impose spatial smoothness. Validation studies on 13 clinical and 5 synthetic brain tumors demonstrated the proposed algorithm outperforms graph cut and grow cut algorithms in all cases with a lower sensitivity to initialization and tumor type.

Learning-based approaches have become increasingly practical in medical imaging. For a supervised learning strategy, the quality of the trained algorithm (usually a classifier) is heavily dependent on the amount, as well as quality, of the available training data. It is often very time-consuming to obtain the ground truth manual delineations. In this paper, we propose a semi-supervised learning algorithm and show its application to skull stripping in brain MRI. The resulting method takes advantage of existing state-of-the-art systems, such as BET and FreeSurfer, to sample unlabeled data in an agreement-based framework. Using just two labeled and a set of unlabeled MRI scans, a voxel-based random forest classifier is trained to perform the skull stripping. Our system is practical, and it displays significant improvement over supervised approaches, BET and FreeSurfer in two datasets (60 test images).

Radiotherapy planning requires accurate delineations of the critical structures. To avoid manual contouring, atlas-based segmentation can be used to get automatic delineations. However, the results strongly depend on the chosen atlas, especially for the head and neck region where the anatomical variability is high. To address this problem, atlases adapted to the patient’s anatomy may allow for a better registration, and already showed an improvement in segmentation accuracy. However, building such atlases requires the definition of a criterion to select among a database the images that are the most similar to the patient. Moreover, the inter-expert variability of manual contouring may be high, and therefore bias the segmentation if selecting only one image for each region. To tackle these issues, we present an original method to design a piecewise most similar atlas. Given a query image, we propose an efficient criterion to select for each anatomical region the K most similar images among a database by considering local volume variations possibly induced by the tumor. Then, we present a new approach to combine the K images selected for each region into a piecewise most similar template. Our results obtained with 105 CT images of the head and neck show that our method reduces the over-segmentation seen with an average atlas while being robust to inter-expert manual segmentation variability.

We present a fully automatic lung lobe segmentation algorithm that is effective in high resolution computed tomography (CT) datasets in the presence of confounding factors such as incomplete fissures (anatomical structures indicating lobe boundaries), advanced disease states, high body mass index (BMI), and low-dose scanning protocols. In contrast to other algorithms that leverage segmentations of auxiliary structures (esp. vessels and airways), we rely only upon image features indicating fissure locations. We employ a particle system that samples the image domain and provides a set of candidate fissure locations. We follow this stage with

maximum a posteriori

(MAP) estimation to eliminate poor candidates and then perform a post-processing operation to remove remaining noise particles. We then fit a thin plate spline (TPS) interpolating surface to the fissure particles to form the final lung lobe segmentation. Results indicate that our algorithm performs comparably to pulmonologist-generated lung lobe segmentations on a set of challenging cases.

The segmentation of soft tissues in medical images is a challenging problem due to the weak boundary, large deformation and serious mutual influence. We present a novel method incorporating both the shape and appearance information in a 3-D graph-theoretic framework to overcome those difficulties for simultaneous segmentation of prostate and bladder. An arc-weighted graph is constructed corresponding to the initial mesh. Both the boundary and region information is incorporated into the graph with learned intensity distribution, which drives the mesh to the best fit of the image. A shape prior penalty is introduced by adding weighted-arcs in the graph, which maintains the original topology of the model and constraints the flexibility of the mesh. The surface-distance constraints are enforced to avoid the leakage between prostate and bladder. The target surfaces are found by solving a maximum flow problem in low-order polynomial time. Both qualitative and quantitative results on prostate and bladder segmentation were promising, proving the power of our algorithm.

Cortical multiple sclerosis lesions are difficult to detect in magnetic resonance images due to poor contrast with surrounding grey matter, spatial variation in healthy grey matter and partial volume effects. We propose using an observer-independent laminar profile-based parcellation method to detect cortical lesions. Following cortical surface extraction, profiles are extended from the white matter surface to the grey matter surface. The cortex is parcellated according to profile intensity and shape features using a k-means classifier. The method is applied to a high-resolution quantitative magnetic resonance data set from a fixed

post mortem

multiple sclerosis brain, and validated using histology.

Segmentation is a fundamental problem in medical image analysis. The use of prior knowledge is often considered to address the ill-posedness of the process. Such a process consists in bringing all training examples in the same reference pose, and then building statistics. During inference, pose parameters are usually estimated first, and then one seeks a compromise between data-attraction and model-fitness with the prior model. In this paper, we propose a novel higher-order Markov Random Field (MRF) model to encode pose-invariant priors and perform 3D segmentation of challenging data. The approach encodes data support in the singleton terms that are obtained using machine learning, and prior constraints in the higher-order terms. A dual-decomposition-based inference method is used to recover the optimal solution. Promising results on challenging data involving segmentation of tissue classes of the human skeletal muscle demonstrate the potentials of the method.

In this paper we present a Markov random field (MRF) driven region-based active contour model (MaRACel) for medical image segmentation. State-of-the-art region-based active contour (RAC) models assume that every spatial location in the image is statistically independent of the others, thereby ignoring valuable contextual information. To address this shortcoming we incorporate a MRF prior into the AC model, further generalizing Chan & Vese’s (CV) and Rousson and Deriche’s (RD) AC models. This incorporation requires a Markov prior that is consistent with the continuous variational framework characteristic of active contours; consequently, we introduce a continuous analogue to the discrete Potts model. To demonstrate the effectiveness of MaRACel, we compare its performance to those of the CV and RD AC models in the following scenarios: (1) the qualitative segmentation of a cancerous lesion in a breast DCE-MR image and (2) the qualitative and quantitative segmentations of prostatic acini (glands) in 200 histopathology images. Across the 200 prostate needle core biopsy histology images, MaRACel yielded an average sensitivity, specificity, and positive predictive value of 71%,95%,74% with respect to the segmented gland boundaries; the CV and RD models have corresponding values of 19%,81%,20% and 53%,88%,56%, respectively.

Prior to performing a robot-assisted coronary artery bypass grafting procedure, a pre-operative computed tomography scan is used to assess patient candidacy and to identify the location of the target vessel. The surgeon then determines the optimal port locations to ensure proper reach to the target with the robotic instruments, while assuming that the heart does not undergo any significant changes between the pre- and intra-operative stages. However, the peri-operative workflow itself leads to changes in heart position and consequently the intra-operative target vessel location. As such, the pre-operative plan must be adequately updated to adjust the target vessel location to better suit the intra-operative condition. Here we propose a technique to predict the position of the peri-operative target vessel location with ~ 3.5 mm RMS accuracy. We believe this technique will potentially reduce the rate of conversion of robot-assisted procedures to traditional open-chest surgery due to poor planning.

Prostate biopsy is the clinical standard for the diagnosis of prostate cancer, and technologies for 3D guidance to targets and recording of biopsy locations are promising approaches to reducing the need for repeated biopsies. In this study, we use image-based non-rigid registration to quantify prostate deformation during needle insertion and biopsy gun firing, in order to provide information useful to the overall assessment of a TRUS-guided biopsy system’s expected targeting error. We recorded mean tissue displacements of up to 0.4 mm, accounting for 16% of the clinically-motivated maximum desired RMS error of a guidance system.

Cone-beam computed tomography (CBCT) is an important image modality for dental surgery planning, with high resolution images at a relative low radiation dose. In these scans the mandibular canal is hardly visible, this is a problem for implant surgery planning. We use anisotropic diffusion filtering to remove noise and enhance the mandibular canal in CBCT scans. For the diffusion tensor we use hybrid diffusion with a continuous switch (HDCS), suitable for filtering both tubular as planar image structures. We focus in this paper on the diffusion discretization schemes. The standard scheme shows good isotropic filtering behavior but is not rotational invariant, the diffusion scheme of Weickert is rotational invariant but suffers from checkerboard artifacts. We introduce a new scheme, in which we numerically optimize the image derivatives. This scheme is rotational invariant and shows good isotropic filtering properties on both synthetic as real CBCT data.

Flexible, ergonomically enhanced surgical robots have important applications to transluminal endoscopic surgery, for which path-following and dynamic shape conformance are essential. In this paper, kinematic control of a snake robot for motion stabilisation under dynamic active constraints is addressed. The main objective is to enable the robot to track the visual target accurately and steadily on deforming tissue whilst conforming to pre-defined anatomical constraints. The motion tracking can also be augmented with manual control. By taking into account the physical limits in terms of maximum frequency response of the system (manifested as a delay between the input of the manipulator and the movement of the end-effector), we show the importance of visual-motor synchronisation for performing accurate smooth pursuit movements. Detailed user experiments are performed to demonstrate the practical value of the proposed control mechanism.

Additional material can be found at

http://www.doc.ic.ac.uk/~kkwok/MICCAI10_video/MICCAI10_514_v3.avi

.

In the last decade the use of interventional X-ray imaging, especially for fluoroscopy-guided procedures, has increased dramatically. Due to this the radiation exposure of the medical staff has also increased. Although radiation protection measures such as lead vests are used there are still unprotected regions, most notably the hands and the head. Over time these regions can receive significant amounts of radiation. In this paper we propose a system for approximating the radiation exposure of a physician during surgery. The goal is to sensibilize physicians to their radiation exposure and to give them a tool to quickly check it. To this end we use a real-time 3D reconstruction system which builds a 3D-representation of all the objects in the room. The reconstructed 3D-representation of the physician is then tracked over time and at each time step in which the X-Ray source is used the radiation received by each body part is accumulated. To simulate the radiation we use a physics-based simulation package. The physician can review his radiation exposure after the intervention and use the collected radiation information over a longer time period in order to minimize his radiation exposure by adjusting his positioning relative to the X-ray source. The system can also be used as an awareness tool for less experienced physicians.

Recent developments in optical spectroscopic techniques have permitted in vivo, in situ cellular and molecular sensing and imaging to allow for real-time tissue characterization, functional assessment, and intraoperative guidance. The small area sensed by these probes, however, presents unique challenges when attempting to obtain useful tissue information in-vivo due to the need to maintain constant distance or contact with the target, and tissue deformation. In practice, the effective area can be increased by translating the tip of the probe over the tissue surface and generating functional maps of the underlying tissue response. However, achieving such controlled motions under manual guidance is very difficult, particularly since the probe is typically passed down the instrument channel of a flexible endoscope. This paper describes a force adaptive multi-spectral imaging system integrated with an articulated robotic endoscope that allows a constant contact force to be maintained between the probe and the tissue as the robot tip is actuated across complex tissue profiles. Detailed phantom and ex-vivo tissue validation is provided.

We present a novel tracking system for patient head motion inside 3D medical scanners. Currently, the system is targeted at the Siemens High Resolution Research Tomograph (HRRT) PET scanner. Partial face surfaces are reconstructed using a miniaturized structured light system. The reconstructed 3D point clouds are matched to a reference surface using a robust iterative closest point algorithm. A main challenge is the narrow geometry requiring a compact structured light system and an oblique angle of observation. The system is validated using a mannequin head mounted on a rotary stage. We compare the system to a standard optical motion tracker based on a rigid tracking tool. Our system achieves an angular RMSE of 0.11° demonstrating its relevance for motion compensated 3D scan image reconstructions as well as its competitiveness against the standard optical system with an RMSE of 0.08°. Finally, we demonstrate qualitative result on real face motion estimation.

Tissue deformation tracking is an important topic of minimally invasive surgery with applications ranging from intra-operative guidance to augmented reality visualisation. In this paper, we present a technique for visual tracking of irregular structures with an arbitrary degree of connectivity in space. The variational formulation of the proposed method ensures that correlation is maximised between tracked points and their computed new positions while the overall structure shape variation is minimised, thus maintaining spatial coherence of the tracked structure. The proposed method is applied to surgical annotation and tracking in 3D for telementoring and path-planning. The results are validated both on a CT-scanned phantom model and

in vivo

, showing an average alignment error of 1.79 mm (± 0.72 mm).

An accurate and robust method to detect curve structures, such as a vessel branch or a guidewire, is essential for many medical imaging applications. A fully automatic method, although highly desired, is prone to detection errors that are caused by image noise and curve-like artifacts. In this paper, we present a novel method to interactively detect a curve structure in a 2D fluoroscopy image with a minimum requirement of human corrections. In this work, a learning based method is used to detect curve segments. Based on the detected segment candidates, a graph is built to search a curve structure as the best path passing through user interactions. Furthermore, our method introduces a novel hyper-graph based optimization method to allow for imposing geometric constraints during the path searching, and to provide a smooth and quickly converged result. With minimum human interactions involved, the method can provide accurate detection results, and has been used in different applications for guidewire and vessel detections.

Articulating digital dental models is often inaccurate and very time-consuming. This paper presents an automated approach to efficiently articulate digital dental models to maximum intercuspation (MI). There are two steps in our method. The first step is to position the models to an initial position based on dental curves and a point matching algorithm. The second step is to finally position the models to the MI position based on our novel approach of using iterative surface-based minimum distance mapping with collision constraints. Finally, our method was validated using 12 sets of digital dental models. The results showed that using our method the digital dental models can be accurately and effectively articulated to MI position.

We present a new image-based respiratory motion compensation method for coronary roadmapping in fluoroscopic images. A temporal analysis scheme is proposed to identify static structures in the image gradient domain. An extended Lucas-Kanade algorithm involving a weighted sum-of-squared-difference (WSSD) measure is proposed to estimate the soft tissue motion in the presence of static structures. A temporally compositional motion model is used to deal with large image motion incurred by deep breathing. Promising results have been shown in the experiments conducted on clinical data.

In the context of minimally invasive surgery, clinical risks are highly associated with surgeons’ skill in manipulating surgical tools and their knowledge of the closed anatomy. A quantitative surgical skill assessment can reduce faulty procedures and prevent some surgical risks. In this paper focusing on sinus surgery, we present two methods to identify both skill level and task type by recording motion data of surgical tools as well as the surgeon’s eye gaze location on the screen. We generate a total of 14 discrete Hidden Markov Models for seven surgical tasks at both expert and novice levels using a repeated

k

-fold evaluation method. The dataset contains 95 expert and 139 novice trials of surgery over a cadaver. The results reveal two insights: eye-gaze data contains skill related structures; and adding this info to the surgical tool motion data improves skill assessment by 13.2% and 5.3% for expert and novice levels, respectively. The proposed system quantifies surgeon’s skill level with an accuracy of 82.5% and surgical task type of 77.8%.

Vitreoretinal surgeons use 0.5mm diameter instruments to manipulate delicate tissue inside the eye while applying imperceptible forces that can cause damage to the retina. We present a system which robotically regulates user-applied forces to the tissue, to minimize the risk of retinal hemorrhage or tear during membrane peeling, a common task in vitreoretinal surgery. Our research platform is based on a cooperatively controlled microsurgery robot. It integrates a custom micro-force sensing surgical pick, which provides conventional surgical function and real time force information. We report the development of a new phantom, which is used to assess robot control, force feedback methods, and our newly implemented auditory sensory substitution to specifically assist membrane peeling. Our findings show that auditory sensory substitution decreased peeling forces in all tests, and that robotic force scaling with audio feedback is the most promising aid in reducing peeling forces and task completion time.

Additional material can be found at

https://ciis.lcsr.jhu.edu/Audio_Sensory_Substitution

.

In prostate brachytherapy, transrectal ultrasound (TRUS) is used to visualize the anatomy, while implanted seeds can be seen in C-arm fluoroscopy. Intra-operative dosimetry optimization requires reconstruction of the implanted seeds from multiple C-arm fluoroscopy images, which in turn requires estimation of the C-arm poses. We estimate the pose of the C-arm by two-stage registration between the 2D fluoroscopy images to a 3D TRUS volume. As single-view 2D/3D registration tends to yield depth error, we first estimate the depth from multiple 2D fluoro images and input this to a single-view 2D/3D registration. A commercial phantom was implanted with seeds and imaged with TRUS and CT. Ground-truth registration was established between the two by radiographic fiducials. Synthetic ground-truth fluoro images were created from the CT volume and registered to the 3D TRUS. The average rotation and translation errors were 1.0° (STD=2.3°) and 0.7mm (STD=1.9 mm), respectively. In data from a human patient, the average rotation and lateral translation errors were 0.6° (STD=3.0°) and 1.5 mm (STD=2.8 mm), respectively, relative to the ground-truth established by a radiographic fiducial. Fully automated image-based C-arm pose estimation was demonstrated in prostate brachytherapy. Accuracy and robustness was excellent on phantom. Early result in human patient data appears clinically adequate.

Novel robotic technologies utilised in surgery need assessment for their effects on the user as well as on technical performance. In this paper, the evolution in

‘cognitive burden’

across visuomotor learning is quantified using a combination of functional near infrared spectroscopy (fNIRS) and graph theory. The results demonstrate escalating costs within the activated cortical network during the intermediate phase of learning which is manifest as an increase in cognitive burden. This innovative application of graph theory and fNIRS enables the economic evaluation of brain behaviour underpinning task execution and how this may be impacted by novel technology and learning. Consequently, this may shed light on how robotic technologies improve human-machine interaction and augment minimally invasive surgical skills acquisition. This work has significant implications for the development and assessment of emergent robotic technologies at cortical level and in elucidating learning-related plasticity in terms of inter-regional cortical connectivity.

Breathing motion complicates many image-guided interventions working on the thorax or upper abdomen. However, prior knowledge provided by a statistical breathing model, can reduce the uncertainties of organ location. In this paper, a prediction framework for statistical motion modeling is presented and different representations of the dynamic data for motion model building of the lungs are investigated. Evaluation carried out on 4D-CT data sets of 10 patients showed that a displacement vector-based representation can reduce most of the respiratory motion with a prediction error of about 2 mm, when assuming the diaphragm motion to be known.

Tracking implanted markers in the prostate during each radiation treatment delivery provides an accurate approximation of prostate location, which enables the use of higher daily doses with tighter margins of the treatment beams and thus improves the efficiency of the radiotherapy. However, the lack of 3D image data with such a technique prevents calculation of delivered dose as required for adaptive planning. We propose to use a reference statistical shape model generated from the planning image and a deformed version of the reference model fitted to the implanted marker locations during treatment to estimate a regionally dense deformation from the planning space to the treatment space. Our method provides a means of estimating the treatment image by mapping planning image data to treatment space via the deformation field and therefore enables the calculation of dose distributions with marker tracking techniques during each treatment delivery.

Deformable guide-wire tracking in fluoroscopic sequences is a challenging task due to the low signal to noise ratio of the images and the apparent complex motion of the object of interest. Common tracking methods are based on data terms that do not differentiate well between medical tools and anatomic background such as ribs and vertebrae. A data term learned directly from fluoroscopic sequences would be more adapted to the image characteristics and could help to improve tracking. In this work, our contribution is to learn the relationship between features extracted from the original image and the tracking error. By randomly deforming a guide-wire model around its ground truth position in one

single

reference frame, we explore the space spanned by these features. Therefore, a guide-wire motion distribution model is learned to reduce the intrisic dimensionality of this feature space. Random deformations and the corresponding features can be then automatically generated. In a regression approach, the function mapping this space to the tracking error is learned. The resulting data term is integrated into a tracking framework based on a second-order MAP-MRF formulation which is optimized by QPBO moves yielding high-quality tracking results. Experiments conducted on two fluoroscopic sequences show that our approach is a promising alternative for deformable tracking of guide-wires.

Magnetic Resonance Imaging (MRI)-guided robotic interventions for aortic valve repair promise to dramatically reduce time and cost of operations when compared to endoscopically guided (EG) procedures. A challenging issue is real-time and robust tracking of anatomical landmark points. The interventional tool should be constantly adjusted via a closed feedback control loop to avoid harming these points while valve repair is taking place in the beating heart. A Bayesian network of particle filter trackers proves capable to produce real-time, yet robust behavior. The algorithm is extremely flexible and general - more sophisticated behaviors can be produced by simply increasing the cardinality of the tracking network. Experimental results on 16 MRI cine sequences highlight the promise of the method.

Additional material can be found at

http://ourpapers.info/miccai10-mri

.

The standard workflow in many image-guided procedures, preoperative imaging followed by intraoperative registration, can be a challenging process and is not readily adaptable to certain anatomical regions such as the wrist. In this study we present an alternative, consisting of a preoperative registration calibration and intraoperative navigation using 3D cone-beam CT. A custom calibration tool was developed to preoperatively register an optical tracking system to the imaging space of a digital angiographic C-arm. This preoperative registration was then applied to perform direct navigation using intraoperatively acquired images for the purposes of an in-vitro wrist fixation procedure. A validation study was performed to assess the stability of the registration and found that the mean registration error was approximately 0.3 mm. When compared to two conventional techniques, our navigated wrist repair achieved equal or better screw placement, with fewer drilling attempts and no additional radiation exposure to the patient. These studies suggest that preoperative registration coupled with direct navigation using procedure-specific graphical rendering, is potentially a highly accurate and effective means of performing image-guided interventions.

The traditional Hessian-related vessel filters often suffer from the problem of handling non-cylindrical objects. To remedy the shortcoming, we present a shape-tuned strain energy density function to measure vessel likelihood in 3D images. Based on the tensor invariants and stress-strain principle in mechanics, a new shape discriminating and vessel strength measure function is formulated. The synthetical and clinical data experiments verify the performance of our method in enhancing complex vascular structures including branches, bifurcations, and feature details.

Image-based computational fluid dynamics provides great promise for evaluation of vascular devices and assessment of surgical procedures. However, many previous studies employ idealized arterial and device models or patient-specific models with a limited number of cases, since the model construction process is tedious and time-consuming. Moreover, in contrast to retrospective studies from existing image data, there is a pressing need of prospective analysis with the goal of surgical planning. Therefore, it is necessary to construct models with implanted devices in a fast, virtual and interactive fashion. The goal of this paper is to develop new geometric methods to deploy stent grafts virtually to patient-specific models constructed from direct 3D segmentation of medical images. A triangular surface representing vessel lumen boundary is extracted from the segmentation. The diseased portion is then clipped and replaced by the surface of a virtual stent graft following the centerline obtained from the clipped portion. A Y-shape stent graft is employed in case of bifurcated arteries. A method to map a 2D strut pattern on the stent graft is also presented. We demonstrate the application of our methods to quantify wall shear stresses and forces acting on stent grafts in personalized surgical planning for endovascular treatment of thoracic and abdominal aortic aneurysms. Our approach enables prospective model construction and may help to increase its throughput required by routine clinical uses in the future.

Prostate cancer is a major health threat for men. For over five years, the U.S. National Cancer Institute has performed prostate biopsies with a magnetic resonance imaging (MRI)-guided robotic system.

Purpose

: A retrospective evaluation methodology and analysis of the clinical accuracy of this system is reported.

Methods

: Using the pre and post-needle insertion image volumes, a registration algorithm that contains a two-step rigid registration followed by a deformable refinement was developed to capture prostate dislocation during the procedure. The method was validated by using three-dimensional contour overlays of the segmented prostates and the registrations were accurate up to 2 mm.

Results

: It was found that tissue deformation was less of a factor than organ displacement. Out of the 82 biopsies from 21 patients, the mean target displacement, needle placement error, and clinical biopsy error was 5.9 mm, 2.3 mm, and 4 mm, respectively.

Conclusion

: The results suggest that motion compensation for organ displacement should be used to improve targeting accuracy.

In image-guided lung intervention, the electromagnetic (EM) tracked needle can be visualized in a pre-procedural CT by registering the EM tracking and the CT coordinate systems. However, there exist discrepancies between the static pre-procedural CT and the patient due to respiratory motion. This paper proposes an online 4-D CT estimation approach to patient-specific respiratory motion compensation. First, the motion patterns between 4-D CT data and respiratory signals such as fiducials from a number of patients are trained in a template space after image registration. These motion patterns can be used to estimate the patient-specific serial CTs from a static 3-D CT and the real-time respiratory signals of that patient, who do not generally take 4-D CTs. Specifically, the respiratory lung field motion vectors are projected onto the Kernel Principal Component Analysis (K-PCA) space, and a motion estimation model is constructed to estimate the lung field motion from the fiducial motion using the ridge regression method based on the least squares support vector machine (LS-SVM). The algorithm can be performed onsite prior to the intervention to generate the serial CT images according to the respiratory signals in advance, and the estimated CTs can be visualized in real-time during the intervention. In experiments, we evaluated the algorithm using leave-one-out strategy on 30 4-D CT data, and the results showed that the average errors of the lung field surfaces are 1.63mm.

Modeling and analyzing surgeries based on signals that are obtained automatically from the operating room (OR) is a field of recent interest. It can be valuable for analyzing and understanding surgical workflow, for skills evaluation and developing context-aware ORs. In minimally invasive surgery, laparoscopic video is easy to record but it is challenging to extract meaningful information from it. We propose a method that uses additional information about tool usage to perform a dimensionality reduction on image features. Using Canonical Correlation Analysis (CCA) a projection of a high-dimensional image feature space to a low dimensional space is obtained such that semantic information is extracted from the video. To model a surgery based on the signals in the reduced feature space two different statistical models are compared. The capability of segmenting a new surgery into phases only based on the video is evaluated. Dynamic Time Warping which strongly depends on the temporal order in combination with CCA shows the best results.

Additional material can be found at

http://campar.in.tum.de/files/publications/blum2010miccai.video.avi

.

The shift to minimally invasive abdominal surgery has increased reliance on image guidance during surgical procedures. However, these images are most often presented independently, increasing the cognitive workload for the surgeon and potentially increasing procedure time. When warm ischemia of an organ is involved, time is an important factor to consider. To address these limitations, we present a more intuitive visualization that combines images in a common augmented reality environment. In this paper, we assess surgeon performance under the guidance of the conventional visualization system and our fusion system using a phantom study that mimics the tumour resection of partial nephrectomy. The RMS error between the fused images was 2.43mm, which is sufficient for our purposes. A faster planning time for the resection was achieved using our fusion visualization system. This result is a positive step towards decreasing risks associated with long procedure times in minimally invasive abdominal interventions.

Blood flow and tissue velocity can be measured using phase-contrast MRI. In this work, the statistical properties of 4D phase-contrast images are derived, and a novel probabilistic blood flow mapping method based on sequential Monte Carlo sampling is presented. The resulting flow maps visualize and quantify the uncertainty in conventional flow visualization techniques such as streamlines and particle traces.

Purpose

: Accurate, practical, and affordable joint encoding on legacy C-arm fluoroscopes is a major technical challenge. Conventional pose tracking methods, like optical cameras and radiographic fiducials, are hampered by significant shortcomings.

Methods

: We propose to retrofit legacy C-arms with a tilt sensing accelerometer for rotation encoding. Our experimental setup consists of affixing an accelerometer to a full scale C-arm with a webcam as an alternative to X-ray imaging for this feasibility research. Ground-truth C-arm poses were obtained from the webcam that tracked a checkerboard plate. From these we constructed a series of angle and structural correction equations that can properly relate the accelerometer angle readings to C-arm pose during surgery and compensate for systematic structural C-arm deformations, such as sagging and bending.

Results

: Real-time tracking of the primary and secondary angle rotations of the C-arm showed an accuracy and precision of less than 0.5 degrees in the entire range of interest.

We are working towards the development of a robotic hand-held surgical device for laparoscopic interventions that enhances the surgeons’ dexterity. In this paper, the kinematics of the end effector is studied. Different choices of kinematics are compared during an evaluation campaign using a virtual reality simulator to find the optimal one: the Yaw-Roll (YR) kinematics. A proof of concept prototype is made based on the results.

Segmentation of surgical devices in fluoroscopic images and in particular of guide-wires is a valuable element during surgery. In cardiac angioplasty, the problem is particularly challenging due to the following reasons: (i) low signal to noise ratio, (ii) the use of 2D images that accumulate information from the whole volume, and (iii) the similarity between the structure of interest and adjacent anatomical structures. In this paper we propose a novel approach to address these challenges, that combines efficiently low-level detection using machine learning techniques, local unsupervised clustering detections and finally high-level perceptual organization of these segments towards its complete reconstruction. The latter handles miss-detections and is based on a local search algorithm. Very promising results were obtained.

We have developed an algorithm for real-time volumetric image reconstruction and 3D tumor localization based on a single x-ray projection image. We first parameterize the deformation vector fields (DVF) of lung motion by principal component analysis (PCA). Then we optimize the DVF applied to a reference image by adapting the PCA coefficients such that the simulated projection of the reconstructed image matches the measured projection. The algorithm was tested on a digital phantom as well as patient data. The average relative image reconstruction error and 3D tumor localization error for the phantom is 7.5% and 0.9 mm, respectively. The tumor localization error for patient is ~2 mm. The computation time of reconstructing one volumetric image from each projection is around 0.2 and 0.3 seconds for phantom and patient, respectively, on an NVIDIA C1060 GPU. Clinical application can potentially lead to accurate 3D tumor tracking from a single imager.

We present a new preoperative planning method for reducing the risk associated with insertion of straight tools in image-guided keyhole neurosurgery. The method quantifies the risks of multiple candidate trajectories and presents them on the outer head surface to assist the neurosurgeon in selecting the safest path. The surgeon can then define and/or revise the trajectory, add a new one using interactive 3D visualization, and obtain a quantitative risk measures. The trajectory risk is evaluated based on the tool placement uncertainty, on the proximity of critical brain structures, and on a predefined table of quantitative geometric risk measures. Our results on five targets show a significant reduction in trajectory risk and a shortening of the preoperative planning time as compared to the current routine method.

It has been shown that excessive white light exposure during retinal microsurgery can induce retinal damage. To address this problem, one can illuminate the retina with a device that alternates between white, and less damaging limited-spectrum light. The surgeon is then presented with a fully colored video by recoloring the limited-spectrum light frames, using information from the white-light frames. To obtain accurately colored images, while reducing phototoxicity, we have developed a novel algorithm that monitors the quality of the recolored images and determines when white light may be substituted by limited-spectrum light. We show qualitatively and quantitatively that our system can provide reliable images using a significantly smaller light dose as compared to other state-of-the-art coloring schemes.

Multi-photon fluorescence microscopy (MFM) captures high-resolution anatomical and functional fluorescence image sequences and can be used for the intact brain imaging of small animals. Recently, it has been extended from imaging anesthetized and head-stabilized animals to awake and head-restrained ones for

in vivo

neurological study. In these applications, motion correction is an important pre-processing step since brain pulsation and tiny body movement can cause motion artifacts and prevent stable serial image acquisition at such a high spatial resolution. This paper proposes a speed embedded hidden Markov model (SEHMM) for motion correction in MFM imaging of awake head-restrained mice. The algorithm extends the traditional HMM method by embedding a motion prediction model to better estimate the state transition probability. SEHMM is a line-by-line motion correction algorithm, which is implemented within the in-focal-plane 2-D videos and can operate directly on the motion-distorted imaging data without external signal measurements such as the movement, heartbeat, respiration, or muscular tension. In experiments, we demonstrat that SEHMM is more accurate than traditional HMM using both simulated and real MFM image sequences.

Pre-operative planning in orthopedic surgery is essential to identify the optimal surgical considerations for each patient-specific case. The planning for osteotomy is presently conducted through two-dimensional (2D) radiographs, where the surgeon has to mentally visualize the bone deformity. This is due to direct three-dimensional (3D) imaging modalities such as Computed Tomography (CT) still being restricted to a minority of complex orthopedic procedures. This paper presents a novel 3D bone reconstruct technique, through bi-planar 2D radiographic images. The reconstruction will be pertinent to osteotomy surgical diagnostics and planning. The framework utilizes a generic 3D model of the bone of interest to obtain the anatomical topology information. A 2D non-rigid registration is performed between the projected contours of this generic 3D model and extracted edges of the X-ray image to identify the planar customization required. Subsequently a free-form deformation based manipulation is conducted to customize the overall 3D bone shape.

A novel approach based on the shape space concept is proposed to classify quasi-conformal deformations of 3D models. A new metric on the quotient space of meshes is introduced to capture changes of the curvature at each vertex of a simplicial complex during deformation. Then, the deformation curve is obtained by calculating the geodesic curve connecting two shapes in the shape space manifold. In order to compare the deformations, the deformation curves are first transferred to the same part of shape space. And then, the Multi-Dimensional Scaling method is employed to eliminate the redundant dimensions facilitating easy comparison of the deformations. To evaluate our method, some synthetic datasets and 23 datasets of gated images of the left heart ventricle during one heartbeat have been examined. Our experiments show that the algorithm can effectively classify normal and abnormal left heart ventricle deformations in shape space.

Colonography is an important screening tool for colorectal lesions. This paper presents a method for establishing spatial correspondence between prone and supine inner colon surfaces reconstructed from CT colonography. The method is able to account for the large deformations and torsions of the colon occurring through movement between the two positions. Therefore, we parameterised the two surfaces in order to provide a 2D indexing system over the full length of the colon using the Ricci flow method. This provides the input to a non-rigid B-spline registration in 2D space which establishes a correspondence for each surface point of the colon in prone and supine position. The method was validated on twelve clinical cases and demonstrated promising registration results over the majority of the colon surface.

Additional material can be found at

http://cmic.cs.ucl.ac.uk/staff/mingxing_hu/endoscopic_images/

.

We present a novel surface smoothing framework using the Laplace-Beltrami eigenfunctions. The Green’s function of an isotropic diffusion equation on a manifold is constructed as a linear combination of the Laplace-Beltraimi operator. The Green’s function is then used in constructing heat kernel smoothing. Unlike many previous approaches, diffusion is analytically represented as a series expansion avoiding numerical instability and inaccuracy issues. This proposed framework is illustrated with mandible surfaces, and is compared to a widely used iterative kernel smoothing technique in computational anatomy. The MATLAB source code is freely available at

http://brainimaging.waisman.wisc.edu/~chung/lb

.

This work explores the problem of solving the MR reconstruction problem when the number of K-space samples acquired in a non-Cartesian grid is considerably less than the resolution (number of pixels) of the image. Mathematically this leads to the solution of an under-determined and ill-posed inverse problem. The inverse problem can only be solved when certain additional/prior assumption is made about the solution. In this case, the prior is the sparsity of the MR image in the wavelet domain. The non-convex lp-norm ( ) of the wavelet coefficient is a suitable metric for sparsity. Such a prior can appear in two forms – in the synthesis prior formulation, the wavelet coefficients of the image is solved for while in the analysis prior formulation the actual image is solved for. Traditionally the synthesis prior formulation is more popular. However, in this work we will show that the analysis prior formulation on redundant wavelet transform provides better MR reconstruction results compared to the synthesis prior formulation.

The task of IMRT planning, particularly in head-and-neck cancer, is a difficult one, often requiring days of work from a trained dosimetrist. One of the main challenges is the prescription of achievable target doses that will be used to optimize a treatment plan. This work explores a data-driven approach in which effort spent on past plans is used to assist in the planning of new patients. Using a database of treated patients, we identify the features of patient geometry that are correlated with received dose and use these to prescribe target dose levels for new patients. We incorporate our approach in a quality-control system, identifying patients with organs that received a dose significantly higher than the one recommended by our method. For all these patients, we have found that a replan using our predicted dose results in noticeable sparing of the organ without compromising dose to other treatment volumes.

The purpose of this study is to characterize the neuroanatomical variations observed in neurological disorders such as dementia. We do a global statistical analysis of brain anatomy and identify the relevant shape deformation patterns that explain corresponding variations in clinical neuropsychological measures. The motivation is to model the inherent relation between anatomical shape and clinical measures and evaluate its statistical significance. We use Partial Least Squares for the multivariate statistical analysis of the deformation momenta under the Large Deformation Diffeomorphic framework. The statistical methodology extracts pertinent directions in the momenta space and the clinical response space in terms of latent variables. We report the results of this analysis on 313 subjects from the Mild Cognitive Impairment group in the Alzheimer’s Disease Neuroimaging Initiative (ADNI).

Adolescent Idiopathic Scoliosis (AIS) characterized by the 3D spine deformity affects about 4% schoolchildren worldwide. One of the prominent theories of the etiopathogenesis of AIS was proposed to be the poor postural balance control due to the impaired vestibular function. Thus, the morphometry of the vestibular system (VS) is of great importance for studying AIS. The VS is a genus-3 structure situated in the inner ear and consists of three semicircular canals lying perpendicular to each other. The high-genus topology of the surface poses great challenge for shape analysis. In this work, we propose an effective method to analyze shapes of high-genus surfaces by considering their geodesic spectra. The key is to compute the canonical hyperbolic geodesic loops of the surface, using the Ricci flow method. The Fuchsian group generators are then computed which can be used to determine the geodesic spectra. The geodesic spectra effectively measure shape differences between high-genus surfaces up to the hyperbolic isometry. We applied the proposed algorithm to the VS of 12 normal and 15 AIS subjects. Experimental results show the effectiveness of our algorithm and reveal statistical shape difference in the VS between right-thoracic AIS and normal subjects.

We introduce a value-based noise reduction method for Dual-Energy CT applications. It is based on joint intensity statistics estimated from high- and low-energy CT scans of the identical anatomy in order to reduce the noise level in both scans. For a given pair of measurement values, a local gradient ascension algorithm in the probability space is used to provide a noise reduced estimate. As a consequence, two noise reduced images are obtained. It was evaluated with synthetic data in terms of quantitative accuracy and contrast to noise ratio (CNR)-gain. The introduced method allows for reducing patient dose by at least 30% while maintaining the original CNR level. Additionally, the dose reduction potential was shown with a radiological evaluation on real patient data. The method can be combined with state-of-the-art filter-based noise reduction techniques, and makes low-dose Dual-Energy CT possible for the full spectrum of quantitative CT applications.

We propose a shape descriptor for 3D ear impressions, derived from a comprehensive set of anatomical features. Motivated by hearing aid (HA) manufacturing, the selection of the anatomical features is carried out according to their uniqueness and importance in HA design. This leads to a

canonical ear signature

that is highly distinctive and potentially well suited for classification. First, the anatomical features are characterized into

generic

topological and geometric features, namely

concavities, elbows, ridges, peaks

, and

bumps

on the surface of the ear. Fast and robust algorithms are then developed for their detection. This indirect approach ensures the generality of the algorithms with potential applications in biomedicine, biometrics, and reverse engineering.

Several sources of uncertainties in shape boundaries in medical images have motivated the use of probabilistic labeling approaches. Although it is well-known that the sample space for the probabilistic representation of a pixel is the unit simplex, standard techniques of statistical shape analysis (e.g. principal component analysis) have been applied to probabilistic data as if they lie in the unconstrained real Euclidean space. Since these techniques are not constrained to the geometry of the simplex, the statistically feasible data produced end up representing invalid (out of the simplex) shapes. By making use of methods for dealing with what is known as compositional or closed data, we propose a new framework intrinsic to the unit simplex for statistical analysis of probabilistic multi-shape anatomy. In this framework, the isometric log-ratio (ILR) transformation is used to isometrically and bijectively map the simplex to the Euclidean real space, where data are analyzed in the same way as unconstrained data and then back-transformed to the simplex. We demonstrate favorable properties of ILR over existing mappings (e.g. LogOdds). Our results on synthetic and brain data exhibit a more accurate statistical analysis of probabilistic shapes.

Dynamic PET imaging performs sequence of data acquisition in order to provide visualization and quantification of physiological changes in specific tissues and organs. The reconstruction of activity maps is generally the first step in dynamic PET. State space

H

 ∞ 

approaches have been proved to be a robust method for PET image reconstruction where, however, temporal constraints are not considered during the reconstruction process. In addition, the state space strategies for PET image reconstruction have been computationally prohibitive for practical usage because of the need for matrix inversion. In this paper, we present a minimax formulation of the dynamic PET imaging problem where a radioisotope decay model is employed as physics-based temporal constraints on the photon counts. Furthermore, a robust steady state

H

 ∞ 

filter is developed to significantly improve the computational efficiency with minimal loss of accuracy. Experiments are conducted on Monte Carlo simulated image sequences for quantitative analysis and validation.

In this paper we introduce a novel approach for inferring articulated spine models from images. A low-dimensional manifold embedding is created from a training set of prior mesh models to establish the patterns of global shape variations. Local appearance is captured from neighborhoods in the manifold once the overall representation converges. Inference with respect to the manifold and shape parameters is performed using a Markov Random Field (MRF). Singleton and pairwise potentials measure the support from the data and shape coherence from neighboring models respectively, while higher-order cliques encode geometrical modes of variation for local vertebra shape warping. Optimization of model parameters is achieved using efficient linear programming and duality. The resulting model is geometrically intuitive, captures the statistical distribution of the underlying manifold and respects image support in the spatial domain. Experimental results on spinal column geometry estimation from CT demonstrate the approach’s potential.

Image-guided surgery systems have a wide range of applications where the level of accuracy required for each application varies from millimeters to low sub-millimeter range. In systems that use optical tracking, it is typical to use point-based registration without any weighting schemes to determine the pose of the tracked tool with very good accuracy. However, recent advancements in methods to estimate the measurement uncertainty for each tracked marker and the development of an anisotropically weighted point-based registration algorithm have allowed for the optical tracking accuracy to be improved. In this article, we demonstrate a new tracking method that improves the tracking accuracy by 20 – 45% over the traditional tracking methodology.

In this paper, a texton-based classification system based on raw pixel representation along with a support vector machine with radial basis function kernel is proposed for the classification of emphysema in computed tomography images of the lung. The proposed approach is tested on 168 annotated regions of interest consisting of normal tissue, centrilobular emphysema, and paraseptal emphysema. The results show the superiority of the proposed approach to common techniques in the literature including moments of the histogram of filter responses based on Gaussian derivatives. The performance of the proposed system, with an accuracy of 96.43%, also slightly improves over a recently proposed approach based on local binary patterns.

The performance of computer-aided diagnosis (CAD) systems can be highly influenced by the training strategy. CAD systems are traditionally trained using available labeled data, extracted from a specific data distribution or from public databases. Due to the wide variability of medical data, these databases might not be representative enough when the CAD system is applied to data extracted from a different clinical setting, diminishing the performance or requiring more labeled samples in order to get better data generalization. In this work, we propose the incorporation of an active learning approach in the training phase of CAD systems for reducing the number of required training samples while maximizing the system performance. The benefit of this approach has been evaluated using a specific CAD system for Diabetic Retinopathy screening. The results show that 1) using a training set obtained from a different data source results in a considerable reduction of the CAD performance; and 2) using active learning the selected training set can be reduced from 1000 to 200 samples while maintaining an area under the Receiver Operating Characteristic curve of 0.856.

We apply sparse Bayesian learning methods, automatic relevance determination (ARD) and predictive ARD (PARD), to Alzheimer’s disease (AD) classification to make accurate prediction and identify critical imaging markers relevant to AD at the same time. ARD is one of the most successful Bayesian feature selection methods. PARD is a powerful Bayesian feature selection method, and provides sparse models that is easy to interpret. PARD selects the model with the best estimate of the predictive performance instead of choosing the one with the largest marginal model likelihood. Comparative study with support vector machine (SVM) shows that ARD/PARD in general outperform SVM in terms of prediction accuracy. Additional comparison with surface-based general linear model (GLM) analysis shows that regions with strongest signals are identified by both GLM and ARD/PARD. While GLM P-map returns significant regions all over the cortex, ARD/PARD provide a small number of relevant and meaningful imaging markers with predictive power, including both cortical and subcortical measures.

A scheme for triaging pulmonary abnormalities in pediatric chest radiographs for specialist interpretation would be useful in resource-poor settings, especially those with a high tuberculosis burden. We assess computer-aided detection of pulmonary pathology in pediatric digital chest X-ray images. The method comprises four phases suggested in the literature: lung field segmentation, lung field subdivision, feature extraction and classification. The output of the system is a probability map for each image, giving an indication of the degree of abnormality of every region in the lung fields; the maps may be used as a visual tool for identifying those cases that need further attention. The system is evaluated on a set of anterior-posterior chest images obtained using a linear slot-scanning digital X-ray machine. The classification results produced an area under the ROC of 0.782, averaged over all regions.

A framework for nodule feature-based extraction is presented to classify lung nodules in low-dose CT slices (LDCT) into four categories: juxta, well-circumscribed, vascularized and pleural-tail, based on the extracted information. The Scale Invariant Feature Transform (SIFT) and an adaptation to Daugman’s Iris Recognition algorithm are used for analysis. The SIFT descriptor results are projected to lower-dimensional subspaces using PCA and LDA. Complex Gabor wavelet nodule response obtained from an adopted Daugman Iris Recognition algorithm revealed improvements from the original Daugman binary iris code. This showed that binarized nodule responses (codes) areinadequate for classification since nodules lack texture concentration as seen in the iris, while the SIFT algorithm projected using PCA showed robustness and precision in classification.

We present a novel approach for extracting cluttered objects based on their morphological properties. Specifically, we address the problem of untangling

Caenorhabditis elegans

clusters in high-throughput screening experiments. We represent the skeleton of each worm cluster by a sparse directed graph whose vertices and edges correspond to worm segments and their adjacencies, respectively. We then search for paths in the graph that are most likely to represent worms while minimizing overlap. The worm likelihood measure is defined on a low-dimensional feature space that captures different worm poses, obtained from a training set of isolated worms. We test the algorithm on 236 microscopy images, each containing 15

C. elegans

worms, and demonstrate successful cluster untangling and high worm detection accuracy.

We address the problem of automatically analyzing lateral cephalometric images as a diagnostic tool for patients suffering from Sleep Disordered Breathing (SDB). First, multiple landmarks and anatomical structures that were previously associated with SDB are localized. Then statistical regression is applied in order to estimate the Respiratory Disturbance Index (RDI), which is the standard measure for the severity of obstructive sleep apnea. The landmark localization employs a new registration method that is based on Local Affine Frames (LAF). Multiple LAFs are sampled per image based on random selection of triplets of keypoints, and are used to register the input image to the training images. The landmarks are then projected from the training images to the query image. Following a refinement step, the tongue, velum and pharyngeal wall are localized. We collected a dataset of 70 images and compare the accuracy of the anatomical landmarks with recent publications, showing preferable performance in localizing most of the anatomical points. Furthermore, we are able to show that the location of the anatomical landmarks and structures predicts the severity of the disorder, obtaining an error of less than 7.5 RDI units for 44% of the patients.

Automatic detection of tuberculosis (TB) on chest radiographs is a difficult problem because of the diverse presentation of the disease. A combination of detection systems for abnormalities and normal anatomy is used to improve detection performance. A textural abnormality detection system operating at the pixel level is combined with a clavicle detection system to suppress false positive responses. The output of a shape abnormality detection system operating at the image level is combined in a next step to further improve performance by reducing false negatives. Strategies for combining systems based on serial and parallel configurations were evaluated using the minimum, maximum, product, and mean probability combination rules. The performance of TB detection increased, as measured using the area under the ROC curve, from 0.67 for the textural abnormality detection system alone to 0.86 when the three systems were combined. The best result was achieved using the sum and product rule in a parallel combination of outputs.

Morphology of anatomical structures can provide important diagnostic information regarding disease. Implicit features of morphology, such as contour smoothness or perimeter-to-area ratio, have been used in the context of computerized decision support classifiers to aid disease diagnosis. These features are usually specific to the domain and application (e.g. margin irregularity is a predictor of malignant breast lesions on DCE-MRI). In this paper we present a framework for extracting Diffeomorphic Based Similarity (DBS) features to capture subtle morphometric differences between shapes that may not be captured by implicit features. Object morphology is represented using the medial axis model and objects are compared by determining correspondences between medial axis models using a cluster-based diffeomorphic registration scheme. To visualize and classify morphometric differences, a manifold learning scheme (Graph Embedding) is employed to identify nonlinear dependencies between medial axis model similarity and calculate DBS. We evaluated our DBS on two clinical problems discriminating: (a) different Gleason grades of prostate cancer using gland morphology on a set of 102 images, and (b) benign and malignant lesions on 44 breast DCE-MRI studies. Precision-recall curves demonstrate DBS features are better able to classify shapes belonging to the same class compared to implicit features. A support vector machine (SVM) classifier is trained to distinguish between different classes utilizing DBS. SVM accuracy was 83 ±4.47 % for distinguishing benign from malignant lesions on breast DCE-MRI and over 80% in distinguishing between intermediate Gleason grades of prostate cancer on digitized histology.

With the wide array of multi scale, multi-modal data now available for disease characterization, the major challenge in integrated disease diagnostics is to able to represent the different data streams in a common framework while overcoming differences in scale and dimensionality. This common knowledge representation framework is an important pre-requisite to develop integrated meta-classifiers for disease classification. In this paper, we present a unified data fusion framework, Semi Supervised Multi Kernel Graph Embedding (SeSMiK-GE). Our method allows for representation of individual data modalities via a combined multi-kernel framework followed by semi- supervised dimensionality reduction, where partial label information is incorporated to embed high dimensional data in a reduced space. In this work we evaluate SeSMiK-GE for distinguishing (a) benign from cancerous (CaP) areas, and (b) aggressive high-grade prostate cancer from indolent low-grade by integrating information from 1.5 Tesla

in vivo

Magnetic Resonance Imaging (anatomic) and Spectroscopy (metabolic). Comparing SeSMiK-GE with unimodal T2w, MRS classifiers and a previous published non-linear dimensionality reduction driven combination scheme (ScEPTre) yielded classification accuracies of (a) 91.3% (SeSMiK), 66.1% (MRI), 82.6% (MRS) and 86.8% (ScEPTre) for distinguishing benign from CaP regions, and (b) 87.5% (SeSMiK), 79.8% (MRI), 83.7% (MRS) and 83.9% (ScEPTre) for distinguishing high and low grade CaP over a total of 19 multi-modal MRI patient studies.