Learning-based deformable registration of MR brain images

Detail-preserving image warping by enforcing smooth image sampling

Multi-phase dynamic contrast-enhanced magnetic resonance imaging image registration makes a substantial contribution to medical image analysis. However, existing methods (e.g., VoxelMorph, CycleMorph) often encounter the problem of image information misalignment in deformable registration tasks, posing challenges to the practical application. To address this issue, we propose a novel smooth image sampling method to align full organic information to realize detail-preserving image warping. In this paper, we clarify that the phenomenon about image information mismatch is attributed to imbalanced sampling. Then, a sampling frequency map constructed by sampling frequency estimators is utilized to instruct smooth sampling by reducing the spatial gradient and discrepancy between all-ones matrix and sampling frequency map. In addition, our estimator determines the sampling frequency of a grid voxel in the moving image by aggregating the sum of interpolation weights from warped non-grid sampling points in its vicinity and vectorially constructs sampling frequency map through projection and scatteration. We evaluate the effectiveness of our approach through experiments on two in-house datasets. The results showcase that our method preserves nearly complete details with ideal registration accuracy compared with several state-of-the-art registration methods. Additionally, our method exhibits a statistically significant difference in the regularity of the registration field compared to other methods, at a significance level of p < 0.05. Our code will be released at https://github.com/QingRui-Sha/SFM.

Multiview Feature Learning With Multiatlas-Based Functional Connectivity Networks for MCI Diagnosis

Functional connectivity (FC) networks built from resting-state functional magnetic resonance imaging (rs-fMRI) has shown promising results for the diagnosis of Alzheimer's disease and its prodromal stage, that is, mild cognitive impairment (MCI). FC is usually estimated as a temporal correlation of regional mean rs-fMRI signals between any pair of brain regions, and these regions are traditionally parcellated with a particular brain atlas. Most existing studies have adopted a predefined brain atlas for all subjects. However, the constructed FC networks inevitably ignore the potentially important subject-specific information, particularly, the subject-specific brain parcellation. Similar to the drawback of the "single view" (versus the "multiview" learning) in medical image-based classification, FC networks constructed based on a single atlas may not be sufficient to reveal the underlying complicated differences between normal controls and disease-affected patients due to the potential bias from that particular atlas. In this study, we propose a multiview feature learning method with multiatlas-based FC networks to improve MCI diagnosis. Specifically, a three-step transformation is implemented to generate multiple individually specified atlases from the standard automated anatomical labeling template, from which a set of atlas exemplars is selected. Multiple FC networks are constructed based on these preselected atlas exemplars, providing multiple views of the FC network-based feature representations for each subject. We then devise a multitask learning algorithm for joint feature selection from the constructed multiple FC networks. The selected features are jointly fed into a support vector machine classifier for multiatlas-based MCI diagnosis. Extensive experimental comparisons are carried out between the proposed method and other competing approaches, including the traditional single-atlas-based method. The results indicate that our method significantly improves the MCI classification, demonstrating its promise in the brain connectome-based individualized diagnosis of brain diseases.

Attention-Guided Hybrid Network for Dementia Diagnosis With Structural MR Images

Deep-learning methods (especially convolutional neural networks) using structural magnetic resonance imaging (sMRI) data have been successfully applied to computer-aided diagnosis (CAD) of Alzheimer's disease (AD) and its prodromal stage [i.e., mild cognitive impairment (MCI)]. As it is practically challenging to capture local and subtle disease-associated abnormalities directly from the whole-brain sMRI, most of those deep-learning approaches empirically preselect disease-associated sMRI brain regions for model construction. Considering that such isolated selection of potentially informative brain locations might be suboptimal, very few methods have been proposed to perform disease-associated discriminative region localization and disease diagnosis in a unified deep-learning framework. However, those methods based on task-oriented discriminative localization still suffer from two common limitations, that is: 1) identified brain locations are strictly consistent across all subjects, which ignores the unique anatomical characteristics of each brain and 2) only limited local regions/patches are used for model training, which does not fully utilize the global structural information provided by the whole-brain sMRI. In this article, we propose an attention-guided deep-learning framework to extract multilevel discriminative sMRI features for dementia diagnosis. Specifically, we first design a backbone fully convolutional network to automatically localize the discriminative brain regions in a weakly supervised manner. Using the identified disease-related regions as spatial attention guidance, we further develop a hybrid network to jointly learn and fuse multilevel sMRI features for CAD model construction. Our proposed method was evaluated on three public datasets (i.e., ADNI-1, ADNI-2, and AIBL), showing superior performance compared with several state-of-the-art methods in both tasks of AD diagnosis and MCI conversion prediction.

Deep learning-based detection and segmentation-assisted management of brain metastases

BACKGROUND

Three-dimensional T1 magnetization prepared rapid acquisition gradient echo (3D-T1-MPRAGE) is preferred in detecting brain metastases (BM) among MRI. We developed an automatic deep learning-based detection and segmentation method for BM (named BMDS net) on 3D-T1-MPRAGE images and evaluated its performance.

METHODS

The BMDS net is a cascaded 3D fully convolution network (FCN) to automatically detect and segment BM. In total, 1652 patients with 3D-T1-MPRAGE images from 3 hospitals (n = 1201, 231, and 220, respectively) were retrospectively included. Manual segmentations were obtained by a neuroradiologist and a radiation oncologist in a consensus reading in 3D-T1-MPRAGE images. Sensitivity, specificity, and dice ratio of the segmentation were evaluated. Specificity and sensitivity measure the fractions of relevant segmented voxels. Dice ratio was used to quantitatively measure the overlap between automatic and manual segmentation results. Paired samples t-tests and analysis of variance were employed for statistical analysis.

RESULTS

The BMDS net can detect all BM, providing a detection result with an accuracy of 100%. Automatic segmentations correlated strongly with manual segmentations through 4-fold cross-validation of the dataset with 1201 patients: the sensitivity was 0.96 ± 0.03 (range, 0.84-0.99), the specificity was 0.99 ± 0.0002 (range, 0.99-1.00), and the dice ratio was 0.85 ± 0.08 (range, 0.62-0.95) for total tumor volume. Similar performances on the other 2 datasets also demonstrate the robustness of BMDS net in correctly detecting and segmenting BM in various settings.

CONCLUSIONS

The BMDS net yields accurate detection and segmentation of BM automatically and could assist stereotactic radiotherapy management for diagnosis, therapy planning, and follow-up.

Constructing Connectome Atlas by Graph Laplacian Learning

The recent development of neuroimaging technology and network theory allows us to visualize and characterize the whole-brain functional connectivity in vivo. The importance of conventional structural image atlas widely used in population-based neuroimaging studies has been well verified. Similarly, a "common" brain connectivity map (also called connectome atlas) across individuals can open a new pathway to interpreting disorder-related brain cognition and behaviors. However, the main obstacle of applying the classic image atlas construction approaches to the connectome data is that a regular data structure (such as a grid) in such methods breaks down the intrinsic geometry of the network connectivity derived from the irregular data domain (in the setting of a graph). To tackle this hurdle, we first embed the brain network into a set of graph signals in the Euclidean space via the diffusion mapping technique. Furthermore, we cast the problem of connectome atlas construction into a novel learning-based graph inference model. It can be constructed by iterating the following processes: (1) align all individual brain networks to a common space spanned by the graph spectrum bases of the latent common network, and (2) learn graph Laplacian of the common network that is in consensus with all aligned brain networks. We have evaluated our novel method for connectome atlas construction in comparison with non-learning-based counterparts. Based on experiments using network connectivity data from populations with neurodegenerative and neuropediatric disorders, our approach has demonstrated statistically meaningful improvement over existing methods.

MetricUNet: Synergistic image- and voxel-level learning for precise prostate segmentation via online sampling

Fully convolutional networks (FCNs), including UNet and VNet, are widely-used network architectures for semantic segmentation in recent studies. However, conventional FCN is typically trained by the cross-entropy or Dice loss, which only calculates the error between predictions and ground-truth labels for pixels individually. This often results in non-smooth neighborhoods in the predicted segmentation. This problem becomes more serious in CT prostate segmentation as CT images are usually of low tissue contrast. To address this problem, we propose a two-stage framework, with the first stage to quickly localize the prostate region, and the second stage to precisely segment the prostate by a multi-task UNet architecture. We introduce a novel online metric learning module through voxel-wise sampling in the multi-task network. Therefore, the proposed network has a dual-branch architecture that tackles two tasks: (1) a segmentation sub-network aiming to generate the prostate segmentation, and (2) a voxel-metric learning sub-network aiming to improve the quality of the learned feature space supervised by a metric loss. Specifically, the voxel-metric learning sub-network samples tuples (including triplets and pairs) in voxel-level through the intermediate feature maps. Unlike conventional deep metric learning methods that generate triplets or pairs in image-level before the training phase, our proposed voxel-wise tuples are sampled in an online manner and operated in an end-to-end fashion via multi-task learning. To evaluate the proposed method, we implement extensive experiments on a real CT image dataset consisting 339 patients. The ablation studies show that our method can effectively learn more representative voxel-level features compared with the conventional learning methods with cross-entropy or Dice loss. And the comparisons show that the proposed method outperforms the state-of-the-art methods by a reasonable margin.

High-Order Laplacian Regularized Low-Rank Representation for Multimodal Dementia Diagnosis

Multimodal heterogeneous data, such as structural magnetic resonance imaging (MRI), positron emission tomography (PET), and cerebrospinal fluid (CSF), are effective in improving the performance of automated dementia diagnosis by providing complementary information on degenerated brain disorders, such as Alzheimer's prodromal stage, i.e., mild cognitive impairment. Effectively integrating multimodal data has remained a challenging problem, especially when these heterogeneous data are incomplete due to poor data quality and patient dropout. Besides, multimodal data usually contain noise information caused by different scanners or imaging protocols. The existing methods usually fail to well handle these heterogeneous and noisy multimodal data for automated brain dementia diagnosis. To this end, we propose a high-order Laplacian regularized low-rank representation method for dementia diagnosis using block-wise missing multimodal data. The proposed method was evaluated on 805 subjects (with incomplete MRI, PET, and CSF data) from the real Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort. Experimental results suggest the effectiveness of our method in three tasks of brain disease classification, compared with the state-of-the-art methods.

Dynamic neural circuit disruptions associated with antisocial behaviors

Antisocial behavior (ASB) is believed to have neural substrates; however, the association between ASB and functional brain networks remains unclear. The temporal variability of the functional connectivity (or dynamic FC) derived from resting-state functional MRI has been suggested as a useful metric for studying abnormal behaviors including ASB. This is the first study using low-frequency fluctuations of the dynamic FC to unravel potential system-level neural correlates with ASB. Specifically, we individually associated the dynamic FC patterns with the ASB scores (measured by Antisocial Process Screening Device) of the male offenders (age: 23.29 ± 3.36 years) based on machine learning. Results showed that the dynamic FCs were associated with individual ASB scores. Moreover, we found that it was mainly the inter-network dynamic FCs that were negatively associated with the ASB severity. Three major high-order cognitive functional networks and the sensorimotor network were found to be more associated with ASB. We further found that impaired behavior in the ASB subjects was mainly associated with decreased FC dynamics in these networks, which may explain why ASB subjects usually have impaired executive control and emotional processing functions. Our study shows that temporal variation of the FC could be a promising tool for ASB assessment, treatment, and prevention.

Incomplete multi-modal representation learning for Alzheimer's disease diagnosis

Alzheimers disease (AD) is a complex neurodegenerative disease. Its early diagnosis and treatment have been a major concern of researchers. Currently, the multi-modality data representation learning of this disease is gradually becoming an emerging research field, attracting widespread attention. However, in practice, data from multiple modalities are only partially available, and most of the existing multi-modal learning algorithms can not deal with the incomplete multi-modality data. In this paper, we propose an Auto-Encoder based Multi-View missing data Completion framework (AEMVC) to learn common representations for AD diagnosis. Specifically, we firstly map the original complete view to a latent space using an auto-encoder network framework. Then, the latent representations measuring statistical dependence learned from the complete view are used to complement the kernel matrix of the incomplete view in the kernel space. Meanwhile, the structural information of original data and the inherent association between views are maintained by graph regularization and Hilbert-Schmidt Independence Criterion (HSIC) constraints. Finally, a kernel based multi-view method is applied to the learned kernel matrix for the acquisition of common representations. Experimental results achieved on Alzheimers Disease Neuroimaging Initiative (ADNI) datasets validate the effectiveness of the proposed method.

Multi-Regression based supervised sample selection for predicting baby connectome evolution trajectory from neonatal timepoint

The connectional map of the baby brain undergoes dramatic changes over the first year of postnatal development, which makes its mapping a challenging task, let alone learning how to predict its evolution. Currently, learning models for predicting brain connectomic developmental trajectories remain broadly absent despite their great potential in spotting atypical neurodevelopmental disorders early. This is most likely due to the scarcity and often incompleteness of longitudinal infant neuroimaging studies for training such models. In this paper, we propose the first approach for progressively predicting longitudinal development of brain networks during the postnatal period solely from a baseline connectome around birth. To this end, a supervised multi-regression sample selection strategy is designed to learn how to identify the best set of neighbors of a testing baseline connectome to eventually predict its evolution trajectory at follow-up timepoints. However, given that the training dataset may have missing samples (connectomes) at certain timepoints, this may affect the training of the predictive model. To overcome this problem, we perform a low-rank tensor completion based on a robust principal component analysis to impute the missing training connectomes by linearly approximating similar complete training networks. In the prediction step, our sample selection strategy aims to preserve spatiotemporal relationships between consecutive timepoints. Therefore, the proposed method learns how to identify the set of the local closest neighbors to a target network by training an ensemble of bidirectional regressors leveraging temporal dependency between consecutive timepoints with a recall to the baseline observations to progressively predict the evolution of a testing network over time. Our method achieves the best prediction results and better captures the dynamic changes of each brain connectome over time in comparison to its ablated versions using leave-one-out cross-validation strategy.

Learning-Based Computer-Aided Prescription Model for Parkinson's Disease: A Data-Driven Perspective

In this article, we study a novel problem: "automatic prescription recommendation for PD patients." To realize this goal, we first build a dataset by collecting 1) symptoms of PD patients, and 2) their prescription drug provided by neurologists. Then, we build a novel computer-aided prescription model by learning the relation between observed symptoms and prescription drug. Finally, for the new coming patients, we could recommend (predict) suitable prescription drug on their observed symptoms by our prescription model. From the methodology part, our proposed model, namely Prescription viA Learning lAtent Symptoms (PALAS), could recommend prescription using the multi-modality representation of the data. In PALAS, a latent symptom space is learned to better model the relationship between symptoms and prescription drug, as there is a large semantic gap between them. Moreover, we present an efficient alternating optimization method for PALAS. We evaluated our method using the data collected from 136 PD patients at Nanjing Brain Hospital, which can be regarded as a large dataset in PD research community. The experimental results demonstrate the effectiveness and clinical potential of our method in this recommendation task, if compared with other competing methods.

Discovering Salient Anatomical Landmarks by Predicting Human Gaze

Anatomical landmarks are a crucial prerequisite for many medical imaging tasks. Usually, the set of landmarks for a given task is predefined by experts. The landmark locations for a given image are then annotated manually or via machine learning methods trained on manual annotations. In this paper, in contrast, we present a method to automatically discover and localize anatomical landmarks in medical images. Specifically, we consider landmarks that attract the visual attention of humans, which we term visually salient landmarks. We illustrate the method for fetal neurosonographic images. First, full-length clinical fetal ultrasound scans are recorded with live sonographer gaze-tracking. Next, a convolutional neural network (CNN) is trained to predict the gaze point distribution (saliency map) of the sonographers on scan video frames. The CNN is then used to predict saliency maps of unseen fetal neurosonographic images, and the landmarks are extracted as the local maxima of these saliency maps. Finally, the landmarks are matched across images by clustering the landmark CNN features. We show that the discovered landmarks can be used within affine image registration, with average landmark alignment errors between 4.1% and 10.9% of the fetal head long axis length.

Multi-modal latent space inducing ensemble SVM classifier for early dementia diagnosis with neuroimaging data

Fusing multi-modality data is crucial for accurate identification of brain disorder as different modalities can provide complementary perspectives of complex neurodegenerative disease. However, there are at least four common issues associated with the existing fusion methods. First, many existing fusion methods simply concatenate features from each modality without considering the correlations among different modalities. Second, most existing methods often make prediction based on a single classifier, which might not be able to address the heterogeneity of the Alzheimer's disease (AD) progression. Third, many existing methods often employ feature selection (or reduction) and classifier training in two independent steps, without considering the fact that the two pipelined steps are highly related to each other. Forth, there are missing neuroimaging data for some of the participants (e.g., missing PET data), due to the participants' "no-show" or dropout. In this paper, to address the above issues, we propose an early AD diagnosis framework via novel multi-modality latent space inducing ensemble SVM classifier. Specifically, we first project the neuroimaging data from different modalities into a latent space, and then map the learned latent representations into the label space to learn multiple diversified classifiers. Finally, we obtain the more reliable classification results by using an ensemble strategy. More importantly, we present a Complete Multi-modality Latent Space (CMLS) learning model for complete multi-modality data and also an Incomplete Multi-modality Latent Space (IMLS) learning model for incomplete multi-modality data. Extensive experiments using the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset have demonstrated that our proposed models outperform other state-of-the-art methods.

Surface-constrained volumetric registration for the early developing brain

The T1-weighted and T2-weighted MRI contrasts of the infant brain evolve drastically during the first year of life. This poses significant challenges to inter- and intra-subject registration, which is key to subsequent statistical analyses. Existing registration methods that do not consider temporal contrast changes are ineffective for infant brain MRI data. To address this problem, we present in this paper a method for deformable registration of infant brain MRI. The key advantage of our method is threefold: (i) To deal with appearance changes, registration is performed based on segmented tissue maps instead of image intensity. Segmentation is performed by using an infant-centric algorithm previously developed by our group. (ii) Registration is carried out with respect to both cortical surfaces and volumetric tissue maps, thus allowing precise alignment of both cortical and subcortical structures. (iii) A dynamic elasticity model is utilized to allow large non-linear deformation. Experimental results in comparison with well-established registration methods indicate that our method yields superior accuracy in both cortical and subcortical alignment.

Quantitative susceptibility mapping based hybrid feature extraction for diagnosis of Parkinson's disease

•

Radiomics features from QSM data have significant clinical value for the diagnosis of PD.

•

CNN features, which is better than Radiomics Features, are also useful in the diagnosis of PD.

•

The combination of radiomics features and CNN features can enhance the diagnostic accuracy.

Parkinson's disease is the second most common neurodegenerative disease in the elderly after Alzheimer's disease. The aetiology and pathogenesis of Parkinson's disease (PD) are still unclear, but the loss of dopaminergic cells and the excessive iron deposition in the substantia nigra (SN) are associated with the pathophysiology. As an imaging technique that can quantitatively reflect the amount of iron deposition, Quantitative Susceptibility Mapping (QSM) has been shown to be a promising modality for the diagnosis of PD. In the present work, we propose a hybrid feature extraction method for PD diagnosis using QSM images. First, we extract radiomics features from the SN using QSM and employ machine learning algorithms to classify PD and normal controls (NC). This approach allows us to investigate which features are most vulnerable to the effects of the disease. Along with this approach, we propose a Convolutional Neural Network (CNN) based method which can extract different features from the QSM image to further support the diagnosis of PD. Finally, we combine these two types of features and we find that the radiomics features and CNN features are complementary to each other, which helps further improve the classification (diagnostic) performance. We conclude that: (1) radiomics features from QSM data have significant clinical value for the diagnosis of PD; (2) CNN features are also useful in the diagnosis of PD; and (3) the combination of radiomics features and CNN features can enhance the diagnostic accuracy.

Region-adaptive Deformable Registration of CT/MRI Pelvic Images via Learning-based Image Synthesis

Registration of pelvic CT and MRI is highly desired as it can facilitate effective fusion of two modalities for prostate cancer radiation therapy, i.e., using CT for dose planning and MRI for accurate organ delineation. However, due to the large inter-modality appearance gaps and the high shape/appearance variations of pelvic organs, the pelvic CT/MRI registration is highly challenging. In this paper, we propose a region-adaptive deformable registration method for multi-modal pelvic image registration. Specifically, to handle the large appearance gaps, we first perform both CT-to-MRI and MRI-to-CT image synthesis by multi-target regression forest (MT-RF). Then, to use the complementary anatomical information in the two modalities for steering the registration, we select key points automatically from both modalities and use them together for guiding correspondence detection in the region-adaptive fashion. That is, we mainly use CT to establish correspondences for bone regions, and use MRI to establish correspondences for soft tissue regions. The number of key points is increased gradually during the registration, to hierarchically guide the symmetric estimation of the deformation fields. Experiments for both intra-subject and inter-subject deformable registration show improved performances compared with state-of-the-art multi-modal registration methods, which demonstrate the potentials of our method to be applied for the routine prostate cancer radiation therapy.

Canonical feature selection for joint regression and multi-class identification in Alzheimer's disease diagnosis

Fusing information from different imaging modalities is crucial for more accurate identification of the brain state because imaging data of different modalities can provide complementary perspectives on the complex nature of brain disorders. However, most existing fusion methods often extract features independently from each modality, and then simply concatenate them into a long vector for classification, without appropriate consideration of the correlation among modalities. In this paper, we propose a novel method to transform the original features from different modalities to a common space, where the transformed features become comparable and easy to find their relation, by canonical correlation analysis. We then perform the sparse multi-task learning for discriminative feature selection by using the canonical features as regressors and penalizing a loss function with a canonical regularizer. In our experiments on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset, we use Magnetic Resonance Imaging (MRI) and Positron Emission Tomography (PET) images to jointly predict clinical scores of Alzheimer's Disease Assessment Scale-Cognitive subscale (ADAS-Cog) and Mini-Mental State Examination (MMSE) and also identify multi-class disease status for Alzheimer's disease diagnosis. The experimental results showed that the proposed canonical feature selection method helped enhance the performance of both clinical score prediction and disease status identification, outperforming the state-of-the-art methods.

Deep Learning in Medical Image Analysis

This review covers computer-assisted analysis of images in the field of medical imaging. Recent advances in machine learning, especially with regard to deep learning, are helping to identify, classify, and quantify patterns in medical images. At the core of these advances is the ability to exploit hierarchical feature representations learned solely from data, instead of features designed by hand according to domain-specific knowledge. Deep learning is rapidly becoming the state of the art, leading to enhanced performance in various medical applications. We introduce the fundamentals of deep learning methods and review their successes in image registration, detection of anatomical and cellular structures, tissue segmentation, computer-aided disease diagnosis and prognosis, and so on. We conclude by discussing research issues and suggesting future directions for further improvement.

Joint prediction of longitudinal development of cortical surfaces and white matter fibers from neonatal MRI

The human brain can be modeled as multiple interrelated shapes (or a multishape), each for characterizing one aspect of the brain, such as the cortex and white matter pathways. Predicting the developing multishape is a very challenging task due to the contrasting nature of the developmental trajectories of the constituent shapes: smooth for the cortical surface and non-smooth for white matter tracts due to changes such as bifurcation. We recently addressed this problem and proposed an approach for predicting the multishape developmental spatiotemporal trajectories of infant brains based only on neonatal MRI data using a set of geometric, dynamic, and fiber-to-surface connectivity features. In this paper, we propose two key innovations to further improve the prediction of multishape evolution. First, for a more accurate cortical surface prediction, instead of simply relying on one neonatal atlas to guide the prediction of the multishape, we propose to use multiple neonatal atlases to build a spatially heterogeneous atlas using the multidirectional varifold representation. This individualizes the atlas by locally maximizing its similarity to the testing baseline cortical shape for each cortical region, thereby better representing the baseline testing cortical surface, which founds the multishape prediction process. Second, for temporally consistent fiber prediction, we propose to reliably estimate spatiotemporal connectivity features using low-rank tensor completion, thereby capturing the variability and richness of the temporal development of fibers. Experimental results confirm that the proposed variants significantly improve the prediction performance of our original multishape prediction framework for both cortical surfaces and fiber tracts shape at 3, 6, and 9 months of age. Our pioneering model will pave the way for learning how to predict the evolution of anatomical shapes with abnormal changes. Ultimately, devising accurate shape evolution prediction models that can help quantify and predict the severity of a brain disorder as it progresses will be of great aid in individualized treatment planning.

Detecting Anatomical Landmarks for Fast Alzheimer's Disease Diagnosis

Structural magnetic resonance imaging (MRI) is a very popular and effective technique used to diagnose Alzheimer's disease (AD). The success of computer-aided diagnosis methods using structural MRI data is largely dependent on the two time-consuming steps: 1) nonlinear registration across subjects, and 2) brain tissue segmentation. To overcome this limitation, we propose a landmark-based feature extraction method that does not require nonlinear registration and tissue segmentation. In the training stage, in order to distinguish AD subjects from healthy controls (HCs), group comparisons, based on local morphological features, are first performed to identify brain regions that have significant group differences. In general, the centers of the identified regions become landmark locations (or AD landmarks for short) capable of differentiating AD subjects from HCs. In the testing stage, using the learned AD landmarks, the corresponding landmarks are detected in a testing image using an efficient technique based on a shape-constrained regression-forest algorithm. To improve detection accuracy, an additional set of salient and consistent landmarks are also identified to guide the AD landmark detection. Based on the identified AD landmarks, morphological features are extracted to train a support vector machine (SVM) classifier that is capable of predicting the AD condition. In the experiments, our method is evaluated on landmark detection and AD classification sequentially. Specifically, the landmark detection error (manually annotated versus automatically detected) of the proposed landmark detector is 2.41 mm , and our landmark-based AD classification accuracy is 83.7%. Lastly, the AD classification performance of our method is comparable to, or even better than, that achieved by existing region-based and voxel-based methods, while the proposed method is approximately 50 times faster.

Scalable High-Performance Image Registration Framework by Unsupervised Deep Feature Representations Learning

Feature selection is a critical step in deformable image registration. In particular, selecting the most discriminative features that accurately and concisely describe complex morphological patterns in image patches improves correspondence detection, which in turn improves image registration accuracy. Furthermore, since more and more imaging modalities are being invented to better identify morphological changes in medical imaging data, the development of deformable image registration method that scales well to new image modalities or new image applications with little to no human intervention would have a significant impact on the medical image analysis community. To address these concerns, a learning-based image registration framework is proposed that uses deep learning to discover compact and highly discriminative features upon observed imaging data. Specifically, the proposed feature selection method uses a convolutional stacked autoencoder to identify intrinsic deep feature representations in image patches. Since deep learning is an unsupervised learning method, no ground truth label knowledge is required. This makes the proposed feature selection method more flexible to new imaging modalities since feature representations can be directly learned from the observed imaging data in a very short amount of time. Using the LONI and ADNI imaging datasets, image registration performance was compared to two existing state-of-the-art deformable image registration methods that use handcrafted features. To demonstrate the scalability of the proposed image registration framework, image registration experiments were conducted on 7.0-T brain MR images. In all experiments, the results showed that the new image registration framework consistently demonstrated more accurate registration results when compared to state of the art.

Scalable High-Performance Image Registration Framework by Unsupervised Deep Feature Representations Learning

Feature selection is a critical step in deformable image registration. In particular, selecting the most discriminative features that accurately and concisely describe complex morphological patterns in image patches improves correspondence detection, which in turn improves image registration accuracy. Furthermore, since more and more imaging modalities are being invented to better identify morphological changes in medical imaging data, the development of deformable image registration method that scales well to new image modalities or new image applications with little to no human intervention would have a significant impact on the medical image analysis community. To address these concerns, a learning-based image registration framework is proposed that uses deep learning to discover compact and highly discriminative features upon observed imaging data. Specifically, the proposed feature selection method uses a convolutional stacked autoencoder to identify intrinsic deep feature representations in image patches. Since deep learning is an unsupervised learning method, no ground truth label knowledge is required. This makes the proposed feature selection method more flexible to new imaging modalities since feature representations can be directly learned from the observed imaging data in a very short amount of time. Using the LONI and ADNI imaging datasets, image registration performance was compared to two existing state-of-the-art deformable image registration methods that use handcrafted features. To demonstrate the scalability of the proposed image registration framework, image registration experiments were conducted on 7.0-T brain MR images. In all experiments, the results showed that the new image registration framework consistently demonstrated more accurate registration results when compared to state of the art.

Multidirectional and Topography-based Dynamic-scale Varifold Representations with Application to Matching Developing Cortical Surfaces

The human cerebral cortex is marked by great complexity as well as substantial dynamic changes during early postnatal development. To obtain a fairly comprehensive picture of its age-induced and/or disorder-related cortical changes, one needs to match cortical surfaces to one another, while maximizing their anatomical alignment. Methods that geodesically shoot surfaces into one another as currents (a distribution of oriented normals) and varifolds (a distribution of non-oriented normals) provide an elegant Riemannian framework for generic surface matching and reliable statistical analysis. However, both conventional current and varifold matching methods have two key limitations. First, they only use the normals of the surface to measure its geometry and guide the warping process, which overlooks the importance of the orientations of the inherently convoluted cortical sulcal and gyral folds. Second, the 'conversion' of a surface into a current or a varifold operates at a fixed scale under which geometric surface details will be neglected, which ignores the dynamic scales of cortical foldings. To overcome these limitations and improve varifold-based cortical surface registration, we propose two different strategies. The first strategy decomposes each cortical surface into its normal and tangent varifold representations, by integrating principal curvature direction field into the varifold matching framework, thus providing rich information of the orientation of cortical folding and better characterization of the complex cortical geometry. The second strategy explores the informative cortical geometric features to perform a dynamic-scale measurement of the cortical surface that depends on the local surface topography (e.g., principal curvature), thereby we introduce the concept of a topography-based dynamic-scale varifold. We tested the proposed varifold variants for registering 12 pairs of dynamically developing cortical surfaces from 0 to 6 months of age. Both variants improved the matching accuracy in terms of closeness to the target surface and the goodness of alignment with regional anatomical boundaries, when compared with three state-of-the-art methods: (1) diffeomorphic spectral matching, (2) conventional current-based surface matching, and (3) conventional varifold-based surface matching.

Predicting infant cortical surface development using a 4D varifold-based learning framework and local topography-based shape morphing

Longitudinal neuroimaging analysis methods have remarkably advanced our understanding of early postnatal brain development. However, learning predictive models to trace forth the evolution trajectories of both normal and abnormal cortical shapes remains broadly absent. To fill this critical gap, we pioneered the first prediction model for longitudinal developing cortical surfaces in infants using a spatiotemporal current-based learning framework solely from the baseline cortical surface. In this paper, we detail this prediction model and even further improve its performance by introducing two key variants. First, we use the varifold metric to overcome the limitations of the current metric for surface registration that was used in our preliminary study. We also extend the conventional varifold-based surface registration model for pairwise registration to a spatiotemporal surface regression model. Second, we propose a morphing process of the baseline surface using its topographic attributes such as normal direction and principal curvature sign. Specifically, our method learns from longitudinal data both the geometric (vertices positions) and dynamic (temporal evolution trajectories) features of the infant cortical surface, comprising a training stage and a prediction stage. In the training stage, we use the proposed varifold-based shape regression model to estimate geodesic cortical shape evolution trajectories for each training subject. We then build an empirical mean spatiotemporal surface atlas. In the prediction stage, given an infant, we select the best learnt features from training subjects to simultaneously predict the cortical surface shapes at all later timepoints, based on similarity metrics between this baseline surface and the learnt baseline population average surface atlas. We used a leave-one-out cross validation method to predict the inner cortical surface shape at 3, 6, 9 and 12 months of age from the baseline cortical surface shape at birth. Our method attained a higher prediction accuracy and better captured the spatiotemporal dynamic change of the highly folded cortical surface than the previous proposed prediction method.

Robust anatomical landmark detection with application to MR brain image registration

Comparison of human brain MR images is often challenged by large inter-subject structural variability. To determine correspondences between MR brain images, most existing methods typically perform a local neighborhood search, based on certain morphological features. They are limited in two aspects: (1) pre-defined morphological features often have limited power in characterizing brain structures, thus leading to inaccurate correspondence detection, and (2) correspondence matching is often restricted within local small neighborhoods and fails to cater to images with large anatomical difference. To address these limitations, we propose a novel method to detect distinctive landmarks for effective correspondence matching. Specifically, we first annotate a group of landmarks in a large set of training MR brain images. Then, we use regression forest to simultaneously learn (1) the optimal sets of features to best characterize each landmark and (2) the non-linear mappings from the local patch appearances of image points to their 3D displacements towards each landmark. The learned regression forests are used as landmark detectors to predict the locations of these landmarks in new images. Because each detector is learned based on features that best distinguish the landmark from other points and also landmark detection is performed in the entire image domain, our method can address the limitations in conventional methods. The deformation field estimated based on the alignment of these detected landmarks can then be used as initialization for image registration. Experimental results show that our method is capable of providing good initialization even for the images with large deformation difference, thus improving registration accuracy.

Automatic classification framework for ventricular septal defects: a pilot study on high-throughput mouse embryo cardiac phenotyping

Intensive international efforts are underway toward phenotyping the entire mouse genome by modifying all its [Formula: see text] genes one-by-one for comparative studies. A workload of this scale has triggered numerous studies harnessing image informatics for the identification of morphological defects. However, existing work in this line primarily rests on abnormality detection via structural volumetrics between wild-type and gene-modified mice, which generally fails when the pathology involves no severe volume changes, such as ventricular septal defects (VSDs) in the heart. Furthermore, in embryo cardiac phenotyping, the lack of relevant work in embryonic heart segmentation, the limited availability of public atlases, and the general requirement of manual labor for the actual phenotype classification after abnormality detection, along with other limitations, have collectively restricted existing practices from meeting the high-throughput demands. This study proposes, to the best of our knowledge, the first fully automatic VSD classification framework in mouse embryo imaging. Our approach leverages a combination of atlas-based segmentation and snake evolution techniques to derive the segmentation of heart ventricles, where VSD classification is achieved by checking whether the left and right ventricles border or overlap with each other. A pilot study has validated our approach at a proof-of-concept level and achieved a classification accuracy of 100% through a series of empirical experiments on a database of 15 images.

Spatial Patterns, Longitudinal Development, and Hemispheric Asymmetries of Cortical Thickness in Infants from Birth to 2 Years of Age

Cortical thickness (CT) is related to normal development and neurodevelopmental disorders. It remains largely unclear how the characteristic patterns of CT evolve in the first 2 years. In this paper, we systematically characterized for the first time the detailed vertex-wise patterns of spatial distribution, longitudinal development, and hemispheric asymmetries of CT at 0, 1, and 2 years of age, via surface-based analysis of 219 longitudinal magnetic resonance images from 73 infants. Despite the dynamic increase of CT in the first year and the little change of CT in the second year, we found that the overall spatial distribution of thin and thick cortices was largely present at birth, and evolved only modestly during the first 2 years. Specifically, the precentral gyrus, postcentral gyrus, occipital cortex, and superior parietal region had thin cortices, whereas the prefrontal, lateral temporal, insula, and inferior parietal regions had thick cortices. We revealed that in the first year thin cortices exhibited low growth rates of CT, whereas thick cortices exhibited high growth rates. We also found that gyri were thicker than sulci, and that the anterior bank of the central sulcus was thicker than the posterior bank. Moreover, we showed rightward hemispheric asymmetries of CT in the lateral temporal and posterior insula regions at birth, which shrank gradually in the first 2 years, and also leftward asymmetries in the medial prefrontal, paracentral, and anterior cingulate cortices, which expanded substantially during this period. This study provides the first comprehensive picture of early patterns and evolution of CT during infancy.

Online updating of context-aware landmark detectors for prostate localization in daily treatment CT images

PURPOSE

In image guided radiation therapy, it is crucial to fast and accurately localize the prostate in the daily treatment images. To this end, the authors propose an online update scheme for landmark-guided prostate segmentation, which can fully exploit valuable patient-specific information contained in the previous treatment images and can achieve improved performance in landmark detection and prostate segmentation.

METHODS

To localize the prostate in the daily treatment images, the authors first automatically detect six anatomical landmarks on the prostate boundary by adopting a context-aware landmark detection method. Specifically, in this method, a two-layer regression forest is trained as a detector for each target landmark. Once all the newly detected landmarks from new treatment images are reviewed or adjusted (if necessary) by clinicians, they are further included into the training pool as new patient-specific information to update all the two-layer regression forests for the next treatment day. As more and more treatment images of the current patient are acquired, the two-layer regression forests can be continually updated by incorporating the patient-specific information into the training procedure. After all target landmarks are detected, a multiatlas random sample consensus (multiatlas RANSAC) method is used to segment the entire prostate by fusing multiple previously segmented prostates of the current patient after they are aligned to the current treatment image. Subsequently, the segmented prostate of the current treatment image is again reviewed (or even adjusted if needed) by clinicians before including it as a new shape example into the prostate shape dataset for helping localize the entire prostate in the next treatment image.

RESULTS

The experimental results on 330 images of 24 patients show the effectiveness of the authors' proposed online update scheme in improving the accuracies of both landmark detection and prostate segmentation. Besides, compared to the other state-of-the-art prostate segmentation methods, the authors' method achieves the best performance.

CONCLUSIONS

By appropriate use of valuable patient-specific information contained in the previous treatment images, the authors' proposed online update scheme can obtain satisfactory results for both landmark detection and prostate segmentation.

Identifying informative imaging biomarkers via tree structured sparse learning for AD diagnosis

Neuroimaging provides a powerful tool to characterize neurodegenerative progression and therapeutic efficacy in Alzheimer's disease (AD) and its prodromal stage-mild cognitive impairment (MCI). However, since the disease pathology might cause different patterns of structural degeneration, which is not pre-known, it is still a challenging problem to identify the relevant imaging markers for facilitating disease interpretation and classification. Recently, sparse learning methods have been investigated in neuroimaging studies for selecting the relevant imaging biomarkers and have achieved very promising results on disease classification. However, in the standard sparse learning method, the spatial structure is often ignored, although it is important for identifying the informative biomarkers. In this paper, a sparse learning method with tree-structured regularization is proposed to capture patterns of pathological degeneration from fine to coarse scale, for helping identify the informative imaging biomarkers to guide the disease classification and interpretation. Specifically, we first develop a new tree construction method based on the hierarchical agglomerative clustering of voxel-wise imaging features in the whole brain, by taking into account their spatial adjacency, feature similarity and discriminability. In this way, the complexity of all possible multi-scale spatial configurations of imaging features can be reduced to a single tree of nested regions. Second, we impose the tree-structured regularization on the sparse learning to capture the imaging structures, and then use them for selecting the most relevant biomarkers. Finally, we train a support vector machine (SVM) classifier with the selected features to make the classification. We have evaluated our proposed method by using the baseline MR images of 830 subjects from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database, which includes 198 AD patients, 167 progressive MCI (pMCI), 236 stable MCI (sMCI), and 229 normal controls (NC). Our experimental results show that our method can achieve accuracies of 90.2 %, 87.2 %, and 70.7 % for classifications of AD vs. NC, pMCI vs. NC, and pMCI vs. sMCI, respectively, demonstrating promising performance compared with other state-of-the-art methods.

Simultaneous and consistent labeling of longitudinal dynamic developing cortical surfaces in infants

The human cerebral cortex develops extremely dynamically in the first 2years of life. Accurate and consistent parcellation of longitudinal dynamic cortical surfaces during this critical stage is essential to understand the early development of cortical structure and function in both normal and high-risk infant brains. However, directly applying the existing methods developed for the cross-sectional studies often generates longitudinally-inconsistent results, thus leading to inaccurate measurements of the cortex development. In this paper, we propose a new method for accurate, consistent, and simultaneous labeling of longitudinal cortical surfaces in the serial infant brain MR images. The proposed method is explicitly formulated as a minimization problem with an energy function that includes a data fitting term, a spatial smoothness term, and a temporal consistency term. Specifically, inspired by multi-atlas based label fusion, the data fitting term is designed to integrate the contributions from multi-atlas surfaces adaptively, according to the similarities of their local cortical folding with that of the subject cortical surface. The spatial smoothness term is then designed to adaptively encourage label smoothness based on the local cortical folding geometries, i.e., allowing label discontinuity at sulcal bottoms (which often are the boundaries of cytoarchitecturally and functionally distinct regions). The temporal consistency term is to adaptively encourage the label consistency among the temporally-corresponding vertices, based on their similarity of local cortical folding. Finally, the entire energy function is efficiently minimized by a graph cuts method. The proposed method has been applied to the parcellation of longitudinal cortical surfaces of 13 healthy infants, each with 6 serial MRI scans acquired at 0, 3, 6, 9, 12 and 18months of age. Qualitative and quantitative evaluations demonstrated both accuracy and longitudinal consistency of the proposed method. By using our method, for the first time, we reveal several hitherto unseen properties of the dynamic and regionally heterogeneous development of the cortical surface area in the first 18months of life.

Hierarchical multi-atlas label fusion with multi-scale feature representation and label-specific patch partition

Multi-atlas patch-based label fusion methods have been successfully used to improve segmentation accuracy in many important medical image analysis applications. In general, to achieve label fusion a single target image is first registered to several atlas images. After registration a label is assigned to each target point in the target image by determining the similarity between the underlying target image patch (centered at the target point) and the aligned image patch in each atlas image. To achieve the highest level of accuracy during the label fusion process it's critical for the chosen patch similarity measurement to accurately capture the tissue/shape appearance of the anatomical structure. One major limitation of existing state-of-the-art label fusion methods is that they often apply a fixed size image patch throughout the entire label fusion procedure. Doing so may severely affect the fidelity of the patch similarity measurement, which in turn may not adequately capture complex tissue appearance patterns expressed by the anatomical structure. To address this limitation, we advance state-of-the-art by adding three new label fusion contributions: First, each image patch is now characterized by a multi-scale feature representation that encodes both local and semi-local image information. Doing so will increase the accuracy of the patch-based similarity measurement. Second, to limit the possibility of the patch-based similarity measurement being wrongly guided by the presence of multiple anatomical structures in the same image patch, each atlas image patch is further partitioned into a set of label-specific partial image patches according to the existing labels. Since image information has now been semantically divided into different patterns, these new label-specific atlas patches make the label fusion process more specific and flexible. Lastly, in order to correct target points that are mislabeled during label fusion, a hierarchical approach is used to improve the label fusion results. In particular, a coarse-to-fine iterative label fusion approach is used that gradually reduces the patch size. To evaluate the accuracy of our label fusion approach, the proposed method was used to segment the hippocampus in the ADNI dataset and 7.0 T MR images, sub-cortical regions in LONI LBPA40 dataset, mid-brain regions in SATA dataset from MICCAI 2013 segmentation challenge, and a set of key internal gray matter structures in IXI dataset. In all experiments, the segmentation results of the proposed hierarchical label fusion method with multi-scale feature representations and label-specific atlas patches are more accurate than several well-known state-of-the-art label fusion methods.

Learning to rank atlases for multiple-atlas segmentation

Recently, multiple-atlas segmentation (MAS) has achieved a great success in the medical imaging area. The key assumption is that multiple atlases have greater chances of correctly labeling a target image than a single atlas. However, the problem of atlas selection still remains unexplored. Traditionally, image similarity is used to select a set of atlases. Unfortunately, this heuristic criterion is not necessarily related to the final segmentation performance. To solve this seemingly simple but critical problem, we propose a learning-based atlas selection method to pick up the best atlases that would lead to a more accurate segmentation. Our main idea is to learn the relationship between the pairwise appearance of observed instances (i.e., a pair of atlas and target images) and their final labeling performance (e.g., using the Dice ratio). In this way, we select the best atlases based on their expected labeling accuracy. Our atlas selection method is general enough to be integrated with any existing MAS method. We show the advantages of our atlas selection method in an extensive experimental evaluation in the ADNI, SATA, IXI, and LONI LPBA40 datasets. As shown in the experiments, our method can boost the performance of three widely used MAS methods, outperforming other learning-based and image-similarity-based atlas selection methods.

Comparative evaluation of registration algorithms in different brain databases with varying difficulty: results and insights

Evaluating various algorithms for the inter-subject registration of brain magnetic resonance images (MRI) is a necessary topic receiving growing attention. Existing studies evaluated image registration algorithms in specific tasks or using specific databases (e.g., only for skull-stripped images, only for single-site images, etc.). Consequently, the choice of registration algorithms seems task- and usage/parameter-dependent. Nevertheless, recent large-scale, often multi-institutional imaging-related studies create the need and raise the question whether some registration algorithms can 1) generally apply to various tasks/databases posing various challenges; 2) perform consistently well, and while doing so, 3) require minimal or ideally no parameter tuning. In seeking answers to this question, we evaluated 12 general-purpose registration algorithms, for their generality, accuracy and robustness. We fixed their parameters at values suggested by algorithm developers as reported in the literature. We tested them in 7 databases/tasks, which present one or more of 4 commonly-encountered challenges: 1) inter-subject anatomical variability in skull-stripped images; 2) intensity homogeneity, noise and large structural differences in raw images; 3) imaging protocol and field-of-view (FOV) differences in multi-site data; and 4) missing correspondences in pathology-bearing images. Totally 7,562 registrations were performed. Registration accuracies were measured by (multi-)expert-annotated landmarks or regions of interest (ROIs). To ensure reproducibility, we used public software tools, public databases (whenever possible), and we fully disclose the parameter settings. We show evaluation results, and discuss the performances in light of algorithms' similarity metrics, transformation models and optimization strategies. We also discuss future directions for the algorithm development and evaluations.

Resolution enhancement of lung 4D-CT via group-sparsity

PURPOSE

4D-CT typically delivers more accurate information about anatomical structures in the lung, over 3D-CT, due to its ability to capture visual information of the lung motion across different respiratory phases. This helps to better determine the dose during radiation therapy for lung cancer. However, a critical concern with 4D-CT that substantially compromises this advantage is the low superior-inferior resolution due to less number of acquired slices, in order to control the CT radiation dose. To address this limitation, the authors propose an approach to reconstruct missing intermediate slices, so as to improve the superior-inferior resolution.

METHODS

In this method the authors exploit the observation that sampling information across respiratory phases in 4D-CT can be complimentary due to lung motion. The authors' approach uses this locally complimentary information across phases in a patch-based sparse-representation framework. Moreover, unlike some recent approaches that treat local patches independently, the authors' approach employs the group-sparsity framework that imposes neighborhood and similarity constraints between patches. This helps in mitigating the trade-off between noise robustness and structure preservation, which is an important consideration in resolution enhancement. The authors discuss the regularizing ability of group-sparsity, which helps in reducing the effect of noise and enables better structural localization and enhancement.

RESULTS

The authors perform extensive experiments on the publicly available DIR-Lab Lung 4D-CT dataset [R. Castillo, E. Castillo, R. Guerra, V. Johnson, T. McPhail, A. Garg, and T. Guerrero, "A framework for evaluation of deformable image registration spatial accuracy using large landmark point sets," Phys. Med. Biol. 54, 1849-1870 (2009)]. First, the authors carry out empirical parametric analysis of some important parameters in their approach. The authors then demonstrate, qualitatively as well as quantitatively, the ability of their approach to achieve more accurate and better localized results over bicubic interpolation as well as a related state-of-the-art approach. The authors also show results on some datasets with tumor, to further emphasize the clinical importance of their method.

CONCLUSIONS

The authors have proposed to improve the superior-inferior resolution of 4D-CT by estimating intermediate slices. The authors' approach exploits neighboring constraints in the group-sparsity framework, toward the goal of achieving better localization and noise robustness. The authors' results are encouraging, and positively demonstrate the role of group-sparsity for 4D-CT resolution enhancement.

Mapping longitudinal development of local cortical gyrification in infants from birth to 2 years of age

Human cortical folding is believed to correlate with cognitive functions. This likely correlation may have something to do with why abnormalities of cortical folding have been found in many neurodevelopmental disorders. However, little is known about how cortical gyrification, the cortical folding process, develops in the first 2 years of life, a period of dynamic and regionally heterogeneous cortex growth. In this article, we show how we developed a novel infant-specific method for mapping longitudinal development of local cortical gyrification in infants. By using this method, via 219 longitudinal 3T magnetic resonance imaging scans from 73 healthy infants, we systemically and quantitatively characterized for the first time the longitudinal cortical global gyrification index (GI) and local GI (LGI) development in the first 2 years of life. We found that the cortical GI had age-related and marked development, with 16.1% increase in the first year and 6.6% increase in the second year. We also found marked and regionally heterogeneous cortical LGI development in the first 2 years of life, with the high-growth regions located in the association cortex, whereas the low-growth regions located in sensorimotor, auditory, and visual cortices. Meanwhile, we also showed that LGI growth in most cortical regions was positively correlated with the brain volume growth, which is particularly significant in the prefrontal cortex in the first year. In addition, we observed gender differences in both cortical GIs and LGIs in the first 2 years, with the males having larger GIs than females at 2 years of age. This study provides valuable information on normal cortical folding development in infancy and early childhood.

Neurodegenerative disease diagnosis using incomplete multi-modality data via matrix shrinkage and completion

In this work, we are interested in predicting the diagnostic statuses of potentially neurodegenerated patients using feature values derived from multi-modality neuroimaging data and biological data, which might be incomplete. Collecting the feature values into a matrix, with each row containing a feature vector of a sample, we propose a framework to predict the corresponding associated multiple target outputs (e.g., diagnosis label and clinical scores) from this feature matrix by performing matrix shrinkage following matrix completion. Specifically, we first combine the feature and target output matrices into a large matrix and then partition this large incomplete matrix into smaller submatrices, each consisting of samples with complete feature values (corresponding to a certain combination of modalities) and target outputs. Treating each target output as the outcome of a prediction task, we apply a 2-step multi-task learning algorithm to select the most discriminative features and samples in each submatrix. Features and samples that are not selected in any of the submatrices are discarded, resulting in a shrunk version of the original large matrix. The missing feature values and unknown target outputs of the shrunk matrix is then completed simultaneously. Experimental results using the ADNI dataset indicate that our proposed framework achieves higher classification accuracy at a greater speed when compared with conventional imputation-based classification methods and also yields competitive performance when compared with the state-of-the-art methods.

Unsupervised deep feature learning for deformable registration of MR brain images

Establishing accurate anatomical correspondences is critical for medical image registration. Although many hand-engineered features have been proposed for correspondence detection in various registration applications, no features are general enough to work well for all image data. Although many learning-based methods have been developed to help selection of best features for guiding correspondence detection across subjects with large anatomical variations, they are often limited by requiring the known correspondences (often presumably estimated by certain registration methods) as the ground truth for training. To address this limitation, we propose using an unsupervised deep learning approach to directly learn the basis filters that can effectively represent all observed image patches. Then, the coefficients by these learnt basis filters in representing the particular image patch can be regarded as the morphological signature for correspondence detection during image registration. Specifically, a stacked two-layer convolutional network is constructed to seek for the hierarchical representations for each image patch, where the high-level features are inferred from the responses of the low-level network. By replacing the hand-engineered features with our learnt data-adaptive features for image registration, we achieve promising registration results, which demonstrates that a general approach can be built to improve image registration by using data-adaptive features through unsupervised deep learning.

Integration of sparse multi-modality representation and anatomical constraint for isointense infant brain MR image segmentation

Segmentation of infant brain MR images is challenging due to poor spatial resolution, severe partial volume effect, and the ongoing maturation and myelination processes. During the first year of life, the brain image contrast between white and gray matters undergoes dramatic changes. In particular, the image contrast inverses around 6-8months of age, where the white and gray matter tissues are isointense in T1 and T2 weighted images and hence exhibit the extremely low tissue contrast, posing significant challenges for automated segmentation. In this paper, we propose a general framework that adopts sparse representation to fuse the multi-modality image information and further incorporate the anatomical constraints for brain tissue segmentation. Specifically, we first derive an initial segmentation from a library of aligned images with ground-truth segmentations by using sparse representation in a patch-based fashion for the multi-modality T1, T2 and FA images. The segmentation result is further iteratively refined by integration of the anatomical constraint. The proposed method was evaluated on 22 infant brain MR images acquired at around 6months of age by using a leave-one-out cross-validation, as well as other 10 unseen testing subjects. Our method achieved a high accuracy for the Dice ratios that measure the volume overlap between automated and manual segmentations, i.e., 0.889±0.008 for white matter and 0.870±0.006 for gray matter.

Incremental learning with selective memory (ILSM): towards fast prostate localization for image guided radiotherapy

Image-guided radiotherapy (IGRT) requires fast and accurate localization of the prostate in 3-D treatment-guided radiotherapy, which is challenging due to low tissue contrast and large anatomical variation across patients. On the other hand, the IGRT workflow involves collecting a series of computed tomography (CT) images from the same patient under treatment. These images contain valuable patient-specific information yet are often neglected by previous works. In this paper, we propose a novel learning framework, namely incremental learning with selective memory (ILSM), to effectively learn the patient-specific appearance characteristics from these patient-specific images. Specifically, starting with a population-based discriminative appearance model, ILSM aims to "personalize" the model to fit patient-specific appearance characteristics. The model is personalized with two steps: backward pruning that discards obsolete population-based knowledge and forward learning that incorporates patient-specific characteristics. By effectively combining the patient-specific characteristics with the general population statistics, the incrementally learned appearance model can localize the prostate of a specific patient much more accurately. This work has three contributions: 1) the proposed incremental learning framework can capture patient-specific characteristics more effectively, compared to traditional learning schemes, such as pure patient-specific learning, population-based learning, and mixture learning with patient-specific and population data; 2) this learning framework does not have any parametric model assumption, hence, allowing the adoption of any discriminative classifier; and 3) using ILSM, we can localize the prostate in treatment CTs accurately (DSC  ∼ 0.89 ) and fast (  ∼ 4 s), which satisfies the real-world clinical requirements of IGRT.

Measuring the dynamic longitudinal cortex development in infants by reconstruction of temporally consistent cortical surfaces

Quantitative measurement of the dynamic longitudinal cortex development during early postnatal stages is of great importance to understand the early cortical structural and functional development. Conventional methods usually reconstruct the cortical surfaces of longitudinal images from the same subject independently, which often generate longitudinally-inconsistent cortical surfaces and thus lead to inaccurate measurement of cortical changes, especially for vertex-wise mapping of cortical development. This paper aims to address this problem by presenting a method to reconstruct temporally-consistent cortical surfaces from longitudinal infant brain MR images, for accurate and consistent measurement of the dynamic cortex development in infants. Specifically, the longitudinal development of the inner cortical surface is first modeled by a deformable growth sheet with elasto-plasticity property to establish longitudinally smooth correspondences of the inner cortical surfaces. Then, the modeled longitudinal inner cortical surfaces are jointly deformed to locate both inner and outer cortical surfaces with a spatial-temporal deformable surface method. The method has been applied to 13 healthy infants, each with 6 serial MR scans acquired at 2 weeks, 3 months, 6 months, 9 months, 12 months and 18 months of age. Experimental results showed that our method with the incorporated longitudinal constraints can reconstruct the longitudinally-dynamic cortical surfaces from serial infant MR images more consistently and accurately than the previously published methods. By using our method, for the first time, we can characterize the vertex-wise longitudinal cortical thickness development trajectory at multiple time points in the first 18 months of life. Specifically, we found the highly age-related and regionally-heterogeneous developmental trajectories of the cortical thickness during this period, with the cortical thickness increased most from 3 to 6 months (16.2%) and least from 9 to 12 months (less than 0.1%). Specifically, the central sulcus only underwent significant increase of cortical thickness from 6 to 9 months and the occipital cortex underwent significant increase from 0 to 9 months, while the frontal, temporal and parietal cortices grew continuously in this first 18 months of life. The adult-like spatial patterns of cortical thickness were generally present at 18 months of age. These results provided detailed insights into the dynamic trajectory of the cortical thickness development in infants.

Latent feature representation with stacked auto-encoder for AD/MCI diagnosis

Recently, there have been great interests for computer-aided diagnosis of Alzheimer's disease (AD) and its prodromal stage, mild cognitive impairment (MCI). Unlike the previous methods that considered simple low-level features such as gray matter tissue volumes from MRI, and mean signal intensities from PET, in this paper, we propose a deep learning-based latent feature representation with a stacked auto-encoder (SAE). We believe that there exist latent non-linear complicated patterns inherent in the low-level features such as relations among features. Combining the latent information with the original features helps build a robust model in AD/MCI classification, with high diagnostic accuracy. Furthermore, thanks to the unsupervised characteristic of the pre-training in deep learning, we can benefit from the target-unrelated samples to initialize parameters of SAE, thus finding optimal parameters in fine-tuning with the target-related samples, and further enhancing the classification performances across four binary classification problems: AD vs. healthy normal control (HC), MCI vs. HC, AD vs. MCI, and MCI converter (MCI-C) vs. MCI non-converter (MCI-NC). In our experiments on ADNI dataset, we validated the effectiveness of the proposed method, showing the accuracies of 98.8, 90.7, 83.7, and 83.3 % for AD/HC, MCI/HC, AD/MCI, and MCI-C/MCI-NC classification, respectively. We believe that deep learning can shed new light on the neuroimaging data analysis, and our work presented the applicability of this method to brain disease diagnosis.