Machine Learning in Medical Imaging

The functional connectome has gained increased attention in the neuroscience community. In general, most network connectivity models are based on correlations between discrete-time series signals that only connect two different brain regions. However, these bivariate region-to-region models do not involve three or more brain regions that form a subnetwork. Here we propose a learning-based method to explore subnetwork biomarkers that are significantly distinguishable between two clinical cohorts. Learning on hypergraph is employed in our work. Specifically, we construct a hypergraph by exhaustively inspecting all possible subnetworks for all subjects, where each hyperedge connects a group of subjects demonstrating highly correlated functional connectivity behavior throughout the underlying subnetwork. The objective function of hypergraph learning is to jointly optimize the weights for all hyperedges which make the separation of two groups by the learned data representation be in the best consensus with the observed clinical labels. We deploy our method to find high order childhood autism biomarkers from rs-fMRI images. Promising results have been obtained from comprehensive evaluation on the discriminative power and generality in diagnosis of Autism.

The registration of abdominal images is an increasing field in research and forms the basis for studying the dynamic motion of organs. Particularly challenging therein are organs which slide along each other. They require discontinuous transform mappings at the sliding boundaries to be accurately aligned. In this paper, we present a novel approach for discontinuity preserving image registration. We base our method on a sparse kernel machine (SKM), where reproducing kernel Hilbert spaces serve as transformation models. We introduce a bilateral regularization term, where neighboring transform parameters are considered jointly. This regularizer enforces a bias to homogeneous regions in the transform mapping and simultaneously preserves discontinuous magnitude changes of parameter vectors pointing in equal directions. We prove a representer theorem for the overall cost function including this bilateral regularizer in order to guarantee a finite dimensional solution. In addition, we build direction-dependent basis functions within the SKM framework in order to elongate the transformations along the potential sliding organ boundaries. In the experiments, we evaluate the registration results of our method on a 4DCT dataset and show superior registration performance of our method over the tested methods.

Approaches for detecting regions of interest in biomedical image data mostly assume that a large amount of annotated training data is available. Certainly, for unchanging problem definitions, the acquisition of large annotated data is time consuming, yet feasible. However, distinct practical problems with large training corpi arise if variability due to different imaging conditions or inter-personal variations lead to significant changes in the image representation. To circumvent these issues, we investigate a classifier learning scenario which requires a small amount of positive annotation data only. Contrarily to previous approaches which focus on methodologies to explicitly or implicitly deal with specific classification scenarios (such as one-class classification), we show that existing supervised classification models can handle a changed setting effectively without any specific modifications.

Two common ways of approaching atlas-based segmentation of brain MRI are (1) intensity-based modelling and (2) multi-atlas label fusion. Intensity-based methods are robust to registration errors but need distinctive image appearances. Multi-atlas label fusion can identify anatomical correspondences with faint appearance cues, but needs a reasonable registration. We propose an ensemble segmentation method that combines the complementary features of both types of approaches. Our method uses the probabilistic estimates of the base methods to compute their optimal combination weights in a spatially varying way. We also propose an intensity-based method (to be used as base method) that offers a trade-off between invariance to registration errors and dependence on distinct appearances. Results show that sacrificing invariance to registration errors (up to a certain degree) improves the performance of our intensity-based method. Our proposed ensemble method outperforms the rest of participating methods in most of the structures of the NeoBrainS12 Challenge on neonatal brain segmentation. We achieve up to $$\sim $$10 % of improvement in some structures.

Quantify dynamic structural changes in the first year of life is a key step in early brain development studies, which is indispensable to accurate deformable image registration. However, very few registration methods can work universally well for infant brain images at arbitrary development stages from birth to one year old, mainly due to (1) large anatomical variations and (2) dynamic appearance changes. In this paper, we propose a novel learning-based registration method to not only align the anatomical structures but also estimate the appearance difference between two infant MR images with possible large age gap. To achieve this goal, we leverage the random forest regression and auto-context model to learn the evolution of shape and appearance from a set of longitudinal infant images (with subject-specific temporal correspondences well determined) and then predict both the deformation pathway and appearance change between two new infant subjects. After that, it becomes much easier to deploy any conventional image registration method to complete the remaining registration since the above challenges for current state-of-the-art registration methods have been solved successfully. We apply our proposed registration method to align infant brain images of different subjects from 2-week-old to 12-month-old. Promising registration results have been achieved in terms of registration accuracy, compared to the counterpart registration methods.

We present a fully automatic system for identifying osteophytes on knee radiographs, and for estimating the widely used Kellgren-Lawrence (KL) grade for Osteoarthritis (OA). We have compared three advanced modelling and texture techniques. We found that a Random Forest trained using Haar-features achieved good results, but the optimal results are obtained by combining shape modelling and texture features. The system achieves the best reported performance for identifying osteophytes (AUC: 0.85), for measuring KL grades and for classifying OA (AUC: 0.93), with an error rate half that of the previous best method.

An effective technique for investigating human brain connectivities, is the reconstruction of fiber orientation distribution functions based on diffusion-weighted MRI. To reconstruct fiber orientations, most current approaches fit a simplified diffusion model, resulting in an approximation error. We present a novel approach for estimating the fiber orientation directly from raw data, by converting the model fitting process into a classification problem based on a convolutional Deep Neural Network, which is able to identify correlated diffusion information within a single voxel. Wevaluate our approach quantitatively on realistic synthetic data as well as on real data and achieve reliable results compared to a state-of-the-art method. This approach is even capable to relieable reconstruct three fiber crossing utilizing only 10 gradient acquisitions.

Quantitative analysis of perivascular spaces (PVSs) is important to reveal the correlations between cerebrovascular lesions and neurodegenerative diseases. In this study, we propose a learning-based segmentation framework to extract the PVSs from high-resolution 7T MR images. Specifically, we integrate three types of vascular filter responses into a structured random forest for classifying voxels into PVS and background. In addition, we also propose a novel entropy-based sampling strategy to extract informative samples in the background for training the classification model. Since various vascular features can be extracted by the three vascular filters, even thin and low-contrast structures can be effectively extracted from the noisy background. Moreover, continuous and smooth segmentation results can be obtained by utilizing the patch-based structured labels. The segmentation performance is evaluated on 19 subjects with 7T MR images, and the experimental results demonstrate that the joint use of entropy-based sampling strategy, vascular features and structured learning improves the segmentation accuracy, with the Dice similarity coefficient reaching 66 %.

The growing collection of longitudinal images for brain disease diagnosis necessitates the development of advanced longitudinal registration and anatomical labeling methods that can respect temporal consistency between images. However, the characteristics of such longitudinal images and how they lodge into the image manifold are often neglected in existing labeling methods. Indeed, most of them independently align atlases to each target time-point image for propagating the pre-defined atlas labels to the subject domain. In this paper, we present a dual-layer groupwise registration method to consistently label anatomical regions of interest in brain images across different time-points using a multi-atlases-based labeling framework. Our framework can best enhance the labeling of longitudinal images through: (1) using the group mean of the longitudinal images of each subject (i.e., subject-mean) as a bridge between atlases and the longitudinal subject scans to align atlases to all time-point images jointly; and (2) using inter-atlas relationship in their nesting manifold to better register each atlas image to the subject-mean. These steps yield to a more consistent (from the joint alignment of atlases with all time-point images) and more accurate (from the manifold-guided registration between each atlases and the subject-mean image) registration, thereby eventually improving the consistency and accuracy for the subsequent labeling step. We have tested our dual-layer groupwise registration method to label two challenging longitudinal brain datasets (i.e., healthy infants and Alzheimer’s disease subjects). Our experimental results have showed that our method achieves higher labeling accuracy while keeping the labeling consistency over time, when compared to the traditional registration scheme (without our proposed contributions). Moreover, the proposed framework can flexibly integrate with the existing label fusion methods, such as sparse-patch based methods, to improve the labeling accuracy of longitudinal datasets.

Neuroimaging data have been widely used to derive possible biomarkers for Alzheimer’s Disease (AD) diagnosis. As only certain brain regions are related to AD progression, many feature selection methods have been proposed to identify informative features (i.e., brain regions) to build an accurate prediction model. These methods mostly only focus on the feature-target relationship to select features which are discriminative to the targets (e.g., diagnosis labels). However, since the brain regions are anatomically and functionally connected, there could be useful intrinsic relationships among features. In this paper, by utilizing both the feature-target and feature-feature relationships, we propose a novel sparse regression model to select informative features which are discriminative to the targets and also representative to the features. We argue that the features which are representative (i.e., can be used to represent many other features) are important, as they signify strong “connection” with other ROIs, and could be related to the disease progression. We use our model to select features for both binary and multi-class classification tasks, and the experimental results on the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset show that the proposed method outperforms other comparison methods considered in this work.

Patch-based segmentation methods utilizing multiple atlases have been widely studied for alleviating some misalignments when registering atlases to the target image. However, weights assigned to the fused labels are typically computed based on predefined features (e.g. simple patch intensities), thus being not necessarily optimal. Due to lack of discriminating features for different regions of an anatomical structure, the original feature space defined by image intensities may limit the segmentation accuracy. To address these problems, we propose a novel local subspace learning based patch-wise label propagation method to estimate a voxel-wise segmentation of the target image. Specifically, multi-scale patch intensities and texture features are first extracted from the image patch in order to acquire the abundant appearance information. Then, margin fisher analysis (MFA) is applied to neighboring samples of each voxel to be segmented from the aligned atlases, in order to extract discriminant features. This process can enhance discrimination of features for different local regions in the anatomical structure. Finally, based on extracted discriminant features, the k-nearest neighbor (kNN) classifier is used to determine the final label for the target voxel. Moreover, for the patch-wise label propagation, we first translate label patches into several discrete class labels by using the k-means clustering method, and then apply MFA to ensure that samples with similar label patches achieve a higher similarity and those with dissimilar label patches achieve a lower similarity. To demonstrate segmentation performance, we comprehensively evaluated the proposed method on the ADNI dataset for hippocampus segmentation. Experimental results show that the proposed method outperforms several conventional multi-atlas based segmentation methods.

In clinical practice, it is more prevalent to use only a single-modal neuroimaging for diagnosis of brain disorders, such as structural magnetic resonance imaging. A neuroimaging dataset generally suffers from the small-sample-size problem, which makes it difficult to train a robust and effective classifier. The learning using privileged information (LUPI) is a newly proposed paradigm, in which the privileged information is available only at the training phase to provide additional information about training samples, but unavailable in the testing phase. LUPI can effectively help construct a better predictive rule to promote classification performance. In this paper, we propose to apply LUPI for the single-modal neuroimaging based diagnosis of brain diseases along with multi-modal training data. Moreover, a boosted LUPI framework is developed, which performs LUPI-based random subspace learning and then ensembles all the LUPI classifiers with the multiple kernel boosting (MKB) algorithm. The experimental results on two neuroimaging datasets show that LUPI-based algorithms are superior to the traditional classifier models for single-modal neuroimaging based diagnosis of brain disorders, and the proposed boosted LUPI framework achieves best performance.

Precise detection and quantification of white matter hyperintensities (WMH) is of great interest in studies of neurodegenerative diseases (NDs). In this work, we propose a novel semi-supervised large margin algorithm for the segmentation of WMH. The proposed algorithm optimizes a kernel based max-margin objective function which aims to maximize the margin averaged over inliers and outliers while exploiting a limited amount of available labelled data. We show that the learning problem can be formulated as a joint framework learning a classifier and a label assignment simultaneously, which can be solved efficiently by an iterative algorithm. We evaluate our method on a database of 280 brain Magnetic Resonance (MR) images from subjects that either suffered from subjective memory complaints or were diagnosed with NDs. The segmented WMH volumes correlate well with the standard clinical measurement (Fazekas score), and both the qualitative visualization results and quantitative correlation scores of the proposed algorithm outperform other well known methods for WMH segmentation.

For neuroimaging-based brain disease diagnosis, sparse regression models have proved their effectiveness in handling high-dimensional data but with a small number of samples. In this paper, we propose a novel framework that utilizes sparse regression models as target-level representation learner and builds a deep convolutional neural network for clinical decision making. Specifically, we first train multiple sparse regression models, each of which has different values of a regularization control parameter, and use the outputs of the trained regression models as target-level representations. Note that sparse regression models trained with different values of a regularization control parameter potentially select different sets of features from the original ones, thereby they have different powers to predict the response values, i.e., a clinical label and clinical scores in our work. We then construct a deep convolutional neural network by taking the target-level representations as input. Our deep network learns to optimally fuse the predicted response variables, i.e., target-level representations, from the same sparse response model(s) and also those from the neighboring sparse response models. To our best knowledge, this is the first work that systematically integrates sparse regression models with deep neural network. In our experiments with ADNI cohort, we validated the effectiveness of the proposed method by achieving the highest classification accuracies in three different tasks of Alzheimer’s disease and mild cognitive impairment identification.

Feature representation is a key step for the classification of histopathological images. The principal component analysis network (PCANet) offers a new unsupervised feature learning algorithm for images via a simple deep network architecture. However, PCA is sensitive to noise and outliers, which may depress the representation learning of PCANet. Grassmann averages (GA) is a newly proposed dimensionality reduction algorithm, which is more robust and effective than PCA. Therefore, in this paper, we propose a GA network (GANet) algorithm to improve the robustness of learned features from images. Moreover, since quaternion algebra provides a mathematically elegant tool to well handle color images, a quaternion representation based GANet (QGANet) is developed to fuse color information and learn a superior representation for color histopathological images. The experimental results on two histopathological image datasets show that GANet outperforms PCANet, while QGANet achieves the best performance for the classification of color histopathological images.

We present a learning-based framework for automatic brain extraction in MR images. It accepts single or multi-contrast brain MR data, builds global binary random forests classifiers at multiple resolution levels, hierarchically performs voxelwise classifications for a test subject, and refines the brain surface using a narrow-band level set technique on the classification map. We further develop a data-driven schema to improve the model performance, which clusters patches of co-registered training images and learns cluster-specific classifiers. We validate our framework via experiments on single and multi-contrast datasets acquired using scanners with different magnetic field strengths. Compared to the state-of-the-art methods, it yields the best performance with statistically significant improvement of the cluster-specific method (with a Dice coefficient of 97.6 ± 0.4 % and an average surface distance of 0.8 ± 0.1 mm) over the global method.

A proof of concept is presented for cross-modality anatomical landmark detection using histograms of unsigned gradient orientations (HUGO) as machine learning image features. This has utility since an existing algorithm trained on data from one modality may be applied to data of a different modality, or data from multiple modalities may be pooled to train one modality-independent algorithm. Landmark detection is performed using a random forest trained on HUGO features and atlas location autocontext features. Three-way cross-modality detection of 20 landmarks is demonstrated in diverse cohorts of CT, MRI T1 and MRI T2 scans of the head. Each cohort is made up of 40 training and 20 test scans, making 180 scans in total. A cross-modality mean landmark error of 5.27 mm is achieved, compared to single-modality error of 4.07 mm.

Holistically detecting interstitial lung disease (ILD) patterns from CT images is challenging yet clinically important. Unfortunately, most existing solutions rely on manually provided regions of interest, limiting their clinical usefulness. In addition, no work has yet focused on predicting more than one ILD from the same CT slice, despite the frequency of such occurrences. To address these limitations, we propose two variations of multi-label deep convolutional neural networks (CNNs). The first uses a deep CNN to detect the presence of multiple ILDs using a regression-based loss function. Our second variant further improves performance, using spatially invariant Fisher Vector encoding of the CNN feature activations. We test our algorithms on a dataset of 533 patients using five-fold cross-validation, achieving high area-under-curve (AUC) scores of 0.982, 0.972, 0.893 and 0.993 for Ground Glass, Reticular, Honeycomb and Emphysema, respectively. As such, our work represents an important step forward in providing clinically effective ILD detection.

Accurate estimation of kidney volume is essential for clinical diagnoses and therapeutic decisions related to renal diseases. Existing kidney volume estimation methods rely on an intermediate segmentation step that is subject to various limitations. In this work, we propose a segmentation-free, supervised learning approach that addresses the challenges of accurate kidney volume estimation caused by extensive variations in kidney shape, size and orientation across subjects. We develop dual regression forests to simultaneously predict the kidney area per image slice, and kidney span per image volume. We validate our method on a dataset of 45 subjects with a total of 90 kidney samples. We obtained a volume estimation accuracy higher than existing segmentation-free (by 72 %) and segmentation-based methods (by 82 %). Compared to a single regression model, the dual regression reduced the false positive area-estimates and improved volume estimation accuracy by 41 %. We also found a mean deviation of under 10 % between our estimated kidney volumes and those obtained manually by expert radiologists.

Correctly classifying a skin lesion is one of the first steps towards treatment. We propose a novel convolutional neural network (CNN) architecture for skin lesion classification designed to learn based on information from multiple image resolutions while leveraging pretrained CNNs. While traditional CNNs are generally trained on a single resolution image, our CNN is composed of multiple tracts, where each tract analyzes the image at a different resolution simultaneously and learns interactions across multiple image resolutions using the same field-of-view. We convert a CNN, pretrained on a single resolution, to work for multi-resolution input. The entire network is fine-tuned in a fully learned end-to-end optimization with auxiliary loss functions. We show how our proposed novel multi-tract network yields higher classification accuracy, outperforming state-of-the-art multi-scale approaches when compared over a public skin lesion dataset.

Retinal image quality assessment (IQA) algorithms use different hand crafted features without considering the important role of the human visual system (HVS). We solve the IQA problem using the principles behind the working of the HVS. Unsupervised information from local saliency maps and supervised information from trained convolutional neural networks (CNNs) are combined to make a final decision on image quality. A novel algorithm is proposed that calculates saliency values for every image pixel at multiple scales to capture global and local image information. This extracts generalized image information in an unsupervised manner while CNNs provide a principled approach to feature learning without the need to define hand-crafted features. The individual classification decisions are fused by weighting them according to their confidence scores. Experimental results on real datasets demonstrate the superior performance of our proposed algorithm over competing methods.

Emphysema is one of the hallmarks of Chronic Obstructive Pulmonary Disorder (COPD), a devastating lung disease often caused by smoking. Emphysema appears on Computed Tomography (CT) scans as a variety of textures that correlate with disease subtypes. It has been shown that the disease subtypes and textures are linked to physiological indicators and prognosis, although neither is well characterized clinically. Most previous computational approaches to modeling emphysema imaging data have focused on supervised classification of lung textures in patches of CT scans. In this work, we describe a generative model that jointly captures heterogeneity of disease subtypes and of the patient population. We also describe a corresponding inference algorithm that simultaneously discovers disease subtypes and population structure in an unsupervised manner. This approach enables us to create image-based descriptors of emphysema beyond those that can be identified through manual labeling of currently defined phenotypes. By applying the resulting algorithm to a large data set, we identify groups of patients and disease subtypes that correlate with distinct physiological indicators.

In many machine learning applications, samples are characterized by a variety of data modalities. In some instances, the training and testing data might include overlapping, but not identical sets of features. In this work, we describe a versatile decision forest methodology to train a classifier based on data that includes several modalities, and then deploy it for use with test data that only presents a subset of the modalities. To this end, we introduce the concept of cross-modality tree feature transforms. These are feature transformations that are guided by how a different feature partitions the training data. We have used the case of staging cognitive impairments to show the benefits of this approach. We train a random forest model that uses both MRI and PET, and can be tested on data that only includes MRI features. We show that the model provides an 8 % improvement in accuracy of separating of progressive cognitive impairments from stable impairments, compared to a model that uses MRI only for training and testing.

In this paper, we present a fully automated machine-learning based solution to localize the fetus and extract the best fetal biometry planes for the head and abdomen from 11–13+6days week 3D fetal ultrasound (US) images. Our method to localize the whole fetus in the sagittal plane utilizes Structured Random Forests (SRFs) and classical Random Forests (RFs). A transfer learning Convolutional Neural Network (CNNs) is then applied to axial images to localize one of three classes (head, body and non-fetal). Finally, the best fetal head and abdomen planes are automatically extracted based on clinical knowledge of the position of the fetal biometry planes within the head and body. Our hybrid method achieves promising localization of the best biometry fetal planes with 1.6 mm and 3.4 mm for head and abdomen plane localization respectively compared to the best manually chosen biometry planes.

Sensorless freehand 3D ultrasound (US) uses speckle decorrelation to estimate small rigid motions between pairs of 2D images. Trajectory estimation combines these motion estimates to obtain the position each image relative to the first. This is prone to the accumulation of measurement bias. Whereas previous work concentrated on correcting biases at the source, this paper proposes to reduce error accumulation by carefully choosing the set of measurements used for trajectory estimation. An undirected graph is created with frames as vertices and motion measurements as edges. Using constrained shortest paths in the graph, random trajectories are generated and averaged to obtain trajectory estimate and uncertainty. To improve accuracy, a Gaussian process regressor is trained on tracked US sequences to predict systematic motion measurement error, which is then used to weigh the edges of the graph. Results on speckle phantom imagery show significantly improved trajectory estimates in comparison with the state-of-the-art, promising accurate volumetric reconstruction.

Brain magnetic resonance image segmentation is one of the most important tasks in medical image analysis and has considerable importance to the effective use of medical imagery in clinical and surgical setting. In particular, the tissue segmentation of white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF) is crucial for brain measurement and disease diagnosis. A variety of studies have shown that the learning-based techniques are efficient and effective in brain tissue segmentation. However, the learning-based segmentation methods depend largely on the availability of good training labels. The commonly used 3T magnetic resonance (MR) images have insufficient image quality and often exhibit poor intensity contrast between WM, GM, and CSF, therefore not able to provide good training labels for learning-based methods. The advances in ultra-high field 7T imaging make it possible to acquire images with an increasingly high level of quality. In this study, we propose an algorithm based on random forest for segmenting 3T MR images by introducing the segmentation information from their corresponding 7T MR images (through semi-automatic labeling). Furthermore, our algorithm iteratively refines the probability maps of WM, GM, and CSF via a cascade of random forest classifiers to improve the tissue segmentation. Experimental results on 10 subjects with both 3T and 7T MR images in a leave-one-out validation, show that the proposed algorithm performs much better than the state-of-the-art segmentation methods.

Resting-state functional magnetic resonance imaging (rs-fMRI) provides key neural imaging characteristics for quantitative assessment and better understanding of the mechanisms of attention deficit hyperactivity disorder (ADHD). Recent multivariate analysis studies showed that functional connectivity (FC) could be used to classify ADHD from normal controls at the individual level. However, there may not be sufficient large numbers of labeled training samples for a hand-on classifier especially for disease classification. In this paper, we propose a transductive maximum margin classification (TMMC) method that uses the available unlabeled data in the learning process. On one hand, the maximum margin classification (MMC) criterion is used to maximize the class margin for the labeled data; on the other hand, a smoothness constraint is imposed on both labeled and unlabeled data projection so that similar samples tend to share the same label. To evaluate the performance of TMMC, experiments on a benchmark cohort from the ADHD-200 competition were performed. The results show that TMMC can improve the performance of ADHD classification using rs-fMRI by involving unlabeled samples, even for small number of labeled training data.

Hippocampal subfields play important and divergent roles in both memory formation and early diagnosis of many neurological diseases, but automatic subfield segmentation is less explored due to its small size and poor image contrast. In this paper, we propose an automatic learning-based hippocampal subfields segmentation framework using multi-modality 3T MR images, including T1 MRI and resting-state fMRI (rs-fMRI). To do this, we first acquire both 3T and 7T T1 MRIs for each training subject, and then the 7T T1 MRI are linearly registered onto the 3T T1 MRI. Six hippocampal subfields are manually labeled on the aligned 7T T1 MRI, which has the 7T image contrast but sits in the 3T T1 space. Next, corresponding appearance and relationship features from both 3T T1 MRI and rs-fMRI are extracted to train a structured random forest as a multi-label classifier to conduct the segmentation. Finally, the subfield segmentation is further refined iteratively by additional context features and updated relationship features. To our knowledge, this is the first work that addresses the challenging automatic hippocampal subfields segmentation using 3T routine T1 MRI and rs-fMRI. The quantitative comparison between our results and manual ground truth demonstrates the effectiveness of our method. Besides, we also find that (a) multi-modality features significantly improved subfield segmentation performance due to the complementary information among modalities; (b) automatic segmentation results using 3T multi-modality images are partially comparable to those on 7T T1 MRI.

This paper proposes a novel method of using regression-guided deformable models for brain regions of interest (ROIs) segmentation. Different from conventional deformable segmentation, which often deforms shape model locally and thus sensitive to initialization, we propose to learn a regressor to explicitly guide the shape deformation, thus eventually improves the performance of ROI segmentation. The regressor is learned via two steps, (1) a joint classification and regression random forest (CRRF) and (2) an auto-context model. The CRRF predicts each voxel’s deformation to the nearest point on the ROI boundary as well as each voxel’s class label (e.g., ROI versus background). The auto-context model further refines all voxel’s deformations (i.e., deformation field) and class labels (i.e., label maps) by considering the neighboring structures. Compared to the conventional random forest regressor, the proposed regressor provides more accurate deformation field estimation and thus more robust in guiding deformation of the shape model. Validated in segmentation of 14 midbrain ROIs from the IXI dataset, our method outperforms the state-of-art multi-atlas label fusion and classification methods, and also significantly reduces the computation cost.

The resting-state functional MRI (rs-fMRI) has been demonstrated as a valuable neuroimaging tool to identify mild cognitive impairment (MCI) patients. Previous studies showed network breakdown in MCI patients with thresholded rs-fMRI connectivity networks. Recently, machine learning techniques have assisted MCI diagnosis by integrating information from multiple networks constructed with a range of thresholds. However, due to the difficulty of searching optimal thresholds, they are often predetermined and uniformly applied to the entire network. Here, we propose an element-wise thresholding strategy to dynamically construct multiple functional networks, i.e., using possibly different thresholds for different elements in the connectivity matrix. These dynamically generated networks are then integrated with a network fusion scheme to capture their common and complementary information. Finally, the features extracted from the fused network are fed into support vector machine (SVM) for MCI diagnosis. Compared to the previous methods, our proposed framework can greatly improve MCI classification performance.

This paper proposes an unsupervised skin lesion segmentation method for dermoscopy images by exploiting the contextual information of skin image at the superpixel level. In particular, a Laplacian sparse coding is presented to evaluate the probabilities of the skin image pixels to delineate lesion border. Moreover, a new rule-based smoothing strategy is proposed as the lesion segmentation refinement procedure. Finally, a multi-scale superpixel segmentation of the skin image is provided to handle size variation of the lesion in order to improve the accuracy of the detected border. Experiments conducted on two datasets show the superiority of our proposed method over several state-of-the-art skin segmentation methods.

Robust and accurate segmentation of the left atrial aneurysm serves as an essential role in the clinical practice. However, automatic segmentation is an extremely challenging task because of the huge shape variabilities of the aneurysm and its complex surroundings. In this paper, we propose a novel framework based on deep residual networks (DRN) for automatic segmentation of the left atrial aneurysm in CT images. Our proposed approach is able to make full use of structure information and adopts extremely deep architectures to learn more discriminative features, which enables more efficient and accurate segmentation. The main procedures of our proposed method are as follows: in the first step, a large-scale of pre-processed images are divided into patches as training units which then are used to train a classification model by DRN; in the second step, based on the trained DRN model, the left atrial aneurysm is segmented with a novel structured fusion algorithm. The proposed method for the first time achieves a fully automatic segmentation of left atrial aneurysm. With sufficient training datasets and test datasets, experimental results show that the proposed framework outperforms the state-of-the-art methods in terms of accuracy and relative error. The proposed method has also a high correlation to the ground truth, which demonstrates it is a promising techniques to left atrial aneurysm segmentation and other clinical applications.

One of the key challenges facing wider adoption of positron emission tomography (PET) as an imaging biomarker of disease is the development of reproducible quantitative image interpretation tools. Quantifying changes in tumor tissue, due to disease progression or treatment regimen, often requires accurate and reproducible delineation of lesions. Lesion segmentation is necessary for measuring tumor proliferation/shrinkage and radiotracer-uptake to quantify tumor metabolism. In this paper, we develop a fully automatic method for lesion delineation, which does not require user-initialization or parameter-tweaking, to segment novel PET images. To achieve this, we train a machine learning system on anatomically and physiologically meaningful imaging cues, to distinguish normal organ activity from tumorous lesion activity. The inferred lesion likelihoods are then used to guide a convex segmentation model, guaranteeing reproducible results. We evaluate our approach on datasets from The Cancer Imaging Archive trained on data from the Quantitative Imaging Network challenge that were delineated by multiple users. Our method not only produces more accurate segmentation than state-of-the art segmentation results, but does so without any user interaction.

Resting-state functional MRI (rfMRI) provides valuable information about functional changes in the brain and is a strong candidate for biomarkers in neurodegenerative diseases. However, commonly used analysis techniques for rfMRI have undesirable features when used for classification. In this paper, we propose a novel supervised learning algorithm based on Linear Discriminant Analysis (LDA) that does not require any decomposition or parcellation of the data and does not need the user to apply any prior knowledge of potential discriminatory networks. Our algorithm extends LDA to obtain a pair of discriminatory spatial maps, and we use computationally efficient methods and regularisation to cope with the large data size, high-dimensionality and low-sample-size typical of rfMRI. The algorithm performs well on simulated rfMRI data, and better than an Independent Component Analysis (ICA)-based discrimination method on a real Parkinson’s disease rfMRI dataset.

Intestinal enteroendocrine cells secrete hormones that are vital for the regulation of glucose metabolism but their differentiation from intestinal stem cells is not fully understood. Asymmetric stem cell divisions have been linked to intestinal stem cell homeostasis and secretory fate commitment. We monitored cell divisions using 4D live cell imaging of cultured intestinal crypts to characterize division modes by means of measurable features such as orientation or shape. A statistical analysis of these measurements requires annotation of mitosis events, which is currently a tedious and time-consuming task that has to be performed manually. To assist data processing, we developed a learning based method to automatically detect mitosis events. The method contains a dual-phase framework for joint detection of dividing cells (mothers) and their progeny (daughters). In the first phase we detect mother and daughters independently using Hough Forest whilst in the second phase we associate mother and daughters by modelling their joint probability as Conditional Random Field (CRF). The method has been evaluated on 32 movies and has achieved an AUC of 72 %, which can be used in conjunction with manual correction and dramatically speed up the processing pipeline.

Although Fourier and Wavelet Transform have been widely used for texture classification methods in medical images, the discrimination performance of FDCT has not been investigated so far in respect to breast cancer detection. Ιn this paper, three multi-resolution transforms, namely the Discrete Wavelet Transform (DWT), the Stationary Wavelet Transform (SWT) and the Fast Discrete Curvelet Transform (FDCT) were comparatively assessed with respect to their ability to discriminate between malignant and benign breast tumors in Dynamic Contrast-Enhanced Magnetic Resonance Images (DCE-MRI). The mean and entropy of the detail sub-images for each decomposition scheme were used as texture features, which were subsequently fed as input into several classifiers. FDCT features fed to a Linear Discriminant Analysis (LDA) classifier produced the highest overall classification performance (93.18 % Accuracy).

For both visual analysis and computer assisted diagnosis systems in breast MRI reading, the delineation and diagnosis of ductal carcinoma in situ (DCIS) is among the most challenging tasks. Recent studies show that kinetic features derived from dynamic contrast enhanced MRI (DCE-MRI) are less effective in discriminating malignant non-masses against benign ones due to their similar kinetic characteristics. Adding shape descriptors can improve the differentiation accuracy. In this work, we propose a set of novel morphological features using the sphere packing technique, aiming to discriminate non-masses based on their shapes. The feature extraction, selection and the classification modules are integrated into a computer-aided diagnosis (CAD) system. The evaluation was performed on a data set of 106 non-masses extracted from 86 patients, which achieved an accuracy of $$90.56\,\%$$, precision of $$90.3\,\%$$, and area under the receiver operating characteristic (ROC) curve (AUC) of 0.94 for the differentiation of benign and malignant types.

This paper proposes a framework of fast neuroimaging-based retrieval and AD analysis, by three key steps: (1) landmark detection, which efficiently extracts landmark-based neuroimaging features without the need of nonlinear registration in testing stage; (2) landmark selection, which removes redundant/noisy landmarks via proposing a feature selection method that considers structural information among landmarks; and (3) hashing, which converts high-dimensional features of subjects into binary codes, for efficiently conducting approximate nearest neighbor search and diagnosis of AD. We have conducted experiments on Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset, and demonstrated that our framework could achieve higher performance than the comparison methods, in terms of accuracy and speed (at least 100 times faster).