Search

Most prior studies focused on developing models for the severity or mortality prediction of COVID-19 patients. However, effective models for recovery-time prediction are still lacking. Here, we present a deep learning solution named iCOVID that can successfully predict the recovery-time of COVID-19 patients based on predefined treatment schemes and heterogeneous multimodal patient information collected within 48 hours after admission. Meanwhile, an interpretable mechanism termed FSR is integrated into iCOVID to reveal the features greatly affecting the prediction of each patient. Data from a total of 3008 patients were collected from three hospitals in Wuhan, China, for large-scale verification. The experiments demonstrate that iCOVID can achieve a time-dependent concordance index of 74.9% (95% CI: 73.6-76.3%) and an average day error of 4.4 days (95% CI: 4.2-4.6 days). Our study reveals that treatment schemes, age, symptoms, comorbidities, and biomarkers are highly related to recovery-time predictions.

Hippocampal atrophy is often considered as one of the important biomarkers for early diagnosis of Alzheimer’s disease (AD), which is an irreversible neurodegenerative disorder. Traditional methods for hippocampus analysis usually computed the shape and volume features from structural Magnetic Resonance Image (sMRI) for the computer-aided diagnosis of AD as well as its prodromal stage, i.e., mild cognitive impairment (MCI). Motivated by the success of deep learning, this paper proposes a deep learning method with the multi-channel cascaded convolutional neural networks (CNNs) to gradually learn the combined hierarchical representations of hippocampal shapes and asymmetries from the binary hippocampal masks for AD classification. First, image segmentation is performed to generate the bilateral hippocampus binary masks for each subject and the mask difference is obtained by subtracting them. Second, multi-channel 3D CNNs are individually constructed on the hippocampus masks and mask differences to extract features of hippocampal shapes and asymmetries for classification. Third, a 2D CNN is cascaded on the 3D CNNs to learn high-level correlation features. Finally, the features learned by multi-channel and cascaded CNNs are combined with a fully connected layer followed by a softmax classifier for disease classification. The proposed method can gradually learn the combined hierarchical features of hippocampal shapes and asymmetries to enhance the classification. Our method is verified on the baseline sMRIs from 807 subjects including 194 AD patients, 397 MCI (164 progressive MCI (pMCI) + 233 stable MCI (sMCI)), and 216 normal controls (NC) from Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset. Experimental results demonstrate that the proposed method achieves an AUC (Area Under the ROC Curve) of 88.4%, 74.6% and 71.9% for AD vs. NC, MCI vs. NC and pMCI vs. sMCI classifications, respectively. It proves the promising classification performance and also shows that both hippocampal shape and asymmetry are helpful for AD diagnosis.

The resting-state functional magnetic resonance imaging (rs-fMRI) reflects functional activity of brain regions by blood-oxygen-level dependent (BOLD) signals. Up to now, many computer-aided diagnosis methods based on rs-fMRI have been developed for Autism Spectrum Disorder (ASD). These methods are mostly the binary classification approaches to determine whether a subject is an ASD patient or not. However, the disease often consists of several sub-categories, which are complex and thus still confusing to many automatic classification methods. Besides, existing methods usually focus on the functional connectivity (FC) features in grey matter regions, which only account for a small portion of the rs-fMRI data. Recently, the possibility to reveal the connectivity information in the white matter regions of rs-fMRI has drawn high attention. To this end, we propose to use the patch-based functional correlation tensor (PBFCT) features extracted from rs-fMRI in white matter, in addition to the traditional FC features from gray matter, to develop a novel multi-class ASD diagnosis method in this work. Our method has two stages. Specifically, in the first stage of multi-source domain adaptation (MSDA), the source subjects belonging to multiple clinical centers (thus called as source domains) are all transformed into the same target feature space. Thus each subject in the target domain can be linearly reconstructed by the transformed subjects. In the second stage of multi-view sparse representation (MVSR), a multi-view classifier for multi-class ASD diagnosis is developed by jointly using both views of the FC and PBFCT features. The experimental results using the ABIDE dataset verify the effectiveness of our method, which is capable of accurately classifying each subject into a respective ASD sub-category.

PURPOSE In clinical practice, invasiveness is an important reference indicator for differentiating the malignant degree of subsolid pulmonary nodules. These nodules can be classified as atypical adenomatous hyperplasia (AAH), adenocarcinoma in situ (AIS), minimally invasive adenocarcinoma (MIA), or invasive adenocarcinoma (IAC). The automatic determination of a nodule's invasiveness based on chest CT scans can guide treatment planning. However, it is challenging, owing to the insufficiency of training data and their interclass similarity and intraclass variation. To address these challenges, we propose a two-stage deep learning strategy for this task: prior-feature learning followed by adaptive-boost deep learning. METHODS The adaptive-boost deep learning is proposed to train a strong classifier for invasiveness classification of subsolid nodules in chest CT images, using multiple 3D convolutional neural network (CNN)-based weak classifiers. Because ensembles of multiple deep 3D CNN models have a huge number of parameters and require large computing resources along with more training and testing time, the prior-feature learning is proposed to reduce the computations by sharing the CNN layers between all weak classifiers. Using this strategy, all weak classifiers can be integrated into a single network. RESULTS Tenfold cross validation of binary classification was conducted on a total of 1357 nodules, including 765 noninvasive (AAH and AIS) and 592 invasive nodules (MIA and IAC). Ablation experimental results indicated that the proposed binary classifier achieved an accuracy of 73.4\%±1.4 with an AUC of 81.3 \%±2.2 . These results are superior compared to those achieved by three experienced chest imaging specialists who achieved an accuracy of 69.1\% , 69.3\% , and 67.9\% , respectively. About 200 additional nodules were also collected. These nodules covered 50 cases for each category (AAH, AIS, MIA, and IAC, respectively). Both binary and multiple classifications were performed on these data and the results demonstrated that the proposed method definitely achieves better performance than the performance achieved by nonensemble deep learning methods. CONCLUSIONS It can be concluded that the proposed adaptive-boost deep learning can significantly improve the performance of invasiveness classification of pulmonary subsolid nodules in CT images, while the prior-feature learning can significantly reduce the total size of deep models. The promising results on clinical data show that the trained models can be used as an effective lung cancer screening tool in hospitals. Moreover, the proposed strategy can be easily extended to other similar classification tasks in 3D medical images.

Handling the data sparseness question is a main way to further enhance the system performances of Head-driven statistical parsing models. Two smoothing methods are proposed to mitigate remaining da...

In this paper, we explore how to decompose the global thermodynamic entropy of a network into components associated with its edges. Commencing from a statistical mechanical picture, in which the normalised Laplacian matrix plays the role of Hamiltonian operator, thermodynamic entropy can be calculated from partition function associated with different energy level occupation distributions arising from Maxwell–Boltzmann statistics. Using the spectral decomposition of the Laplacian, we show how to project the edge-entropy components so that the detailed distribution of entropy across the edges of a network can be achieved. We apply the resulting method to the brain functional connectivity networks using BOLD-fMRI data. The entropic measurement turns out to be an effective tool for the diagnosis of Alzheimer’s disease by finding the most salient functional connectivity features from the corresponding anatomical brain regions.

Endoscopic ultrasonography (EUS) has shown great advantages in the diagnosis and staging of gastrointestinal malignant tumors. However, EUS based diagnosis is limited by variability in the examiner’s subjective interpretation to differentiate between normal and early esophageal carcinoma. In this paper, we propose a novel approach aiming at automatic abnormal region detection from the esophageal EUS images; the contribution is three-fold: first, we present a series of preprocessing strategies developed specifically for the enhancement of EUS images to aid the estimation in the subsequent works. Second, we provide an automatic layer segmentation method based on the multiple surface graph searching approach with incorporation of geometric constraints, which is applied to segment the EUS images into five discernible layers. Third, we introduce the novel feature extraction strategy by utilizing the features from each column in the segmented layers. The SVM classifier is then applied to fulfill the normal and early esophageal carcinoma classification. Subsequently, a clustering method is used to assemble the abnormal columns together so as to detect the abnormal region. Experimental results show that our method has demonstrated its robustness even facing noisy EUS images, and has achieved high accuracy in detecting abnormal region.

Multi-atlas-based methods are commonly used for MR brain image labeling, which alleviates the burdening and time-consuming task of manual labeling in neuroimaging analysis studies. Traditionally, multi-atlas-based methods first register multiple atlases to the target image, and then propagate the labels from the labeled atlases to the unlabeled target image. However, the registration step involves non-rigid alignment, which is often time-consuming and might lack high accuracy. Alternatively, patch-based methods have shown promise in relaxing the demand for accurate registration, but they often require the use of hand-crafted features. Recently, deep learning techniques have demonstrated their effectiveness in image labeling, by automatically learning comprehensive appearance features from training images. In this paper, we propose a multi-atlas guided fully convolutional network (MA-FCN) for automatic image labeling, which aims at further improving the labeling performance with the aid of prior knowledge from the training atlases. Specifically, we train our MA-FCN model in a patch-based manner, where the input data consists of not only a training image patch but also a set of its neighboring (i.e., most similar) affine-aligned atlas patches. The guidance information from neighboring atlas patches can help boost the discriminative ability of the learned FCN. Experimental results on different datasets demonstrate the effectiveness of our proposed method, by significantly outperforming the conventional FCN and several state-of-the-art MR brain labeling methods.

Automatic and accurate pulmonary nodule detection from Computed Tomography (CT) scans plays a vital role in efficient pulmonary cancer screening. Although recent anchor-based methods using Convolutional Neural Networks (CNNs) have achieved state-of-the-art performance in this task, they still have some limitations. First, they do not utilize any prior information such as blood vessel segmentation from images, which can effectively help the detection task. Second, they do not integrate enough context information in the nodule classification branch of the detection framework. Third, the detection is generally achieved by using one single model, which may be insufficient to produce satisfactory results. To overcome these limitations, here we first extract anatomical structures from CT scans and propose a weighted training patch sampling method based on the anatomical structure information. Besides, we propose a parallel multi-task region-of-interest (RoI) head for nodule classification. The proposed method is evaluated on the public Lung Nodule Analysis (LUNA16) challenge dataset. Our method achieves 98.1% in sensitivity at one false-positive per scan and 99.4% in sensitivity at two false-positives per scan.

Segmentation of brain regions for hydrocephalus MR images is pivotally important for quantitatively evaluating patients’ abnormalities. However, the brain image data obtained from hydrocephalus patients always have large deformations and lesion occupancies compared to the normal subjects. This leads to the disruption of the brain’s anatomical structure and the dramatic changes in the shape and location of the brain regions, which poses a significant challenge to the segmentation task. In this paper, we propose a novel segmentation framework, with two modules to better locate and segment these highly distorted brain regions. First, to provide the global anatomical structure information and the absolute position of target regions for segmentation, we use a dual-path registration network which is incorporated into the framework and trained simultaneously together. Second, we develop a novel Positional Correlation Attention Block (PCAB) to introduce the local prior information about the relative positional correlations between different regions, so that the segmentation network can be guided in locating the target regions. In this way, the segmentation framework can be trained with spatial guidance from both global and local positional priors to ensure the robustness of the segmentation. We evaluated our method on the brain MR data of hydrocephalus patients by segmenting 17 consciousness-related ROIs and demonstrated that the proposed method can achieve high performance on the image data with high variations of deformations. Source code is available at: https://github.com/JoeeYF/TBI-Brain-Region-Segmentation.

While convolutional neural networks (CNNs) are dominating the area of computer-aided 3D medical image diagnosis, they are incapable of capturing global information due to the intrinsic locality of convolution. Transformers, another type of neural network empowered with self-attention mechanism, are good at representing global relations, yet computationally expensive and do not generalize well on small datasets. Applying Transformers on 3D medical images has two major problems: 1) medical 3D volumes are bigger in size than natural images which makes training process computationally impractical, 2) and 3D medical image datasets are usually smaller than natural image datasets since medical images are expensive to collect. In this paper, we propose the 3D Medical image Transformer (3DMeT) to address these two issues. 3DMeT introduces 3D convolutional layers to perform block embedding instead of the original linear embedding to cut the computational cost. Additionally, we propose a teacher-student training strategy to address the data-hungry issue by adapting convolutional layers’ weights from a CNN teacher. We conduct experiments on knee images, results demonstrate that the 3DMeT (70.2) confidently outperforms the 3DCNNs (65.3) and Vision Transformer (58.7).

Accurate identification of Mild Cognitive Impairment (MCI) based on resting-state functional Magnetic Resonance Imaging (RS-fMRI) is crucial for reducing the risk of developing Alzheimer’s disease (AD). In the literature, functional connectivity (FC) is often used to extract brain network features. However, it still remains challenging for the estimation of FC because RS-fMRI data are often high-dimensional and small in sample size. Although various Lasso-type sparse learning feature selection methods have been adopted to identify the most discriminative features for brain disease diagnosis, they suffer from two common drawbacks. First, Lasso is instable and not very satisfactory for the high-dimensional and small sample size problem. Second, existing Lasso-type feature selection methods have not simultaneously encapsulate the joint correlations between pairwise features and the target, the correlations between pairwise features, and the joint feature interaction into the feature selection process, thus may lead to suboptimal solutions. To overcome these issues, we propose a novel sparse learning feature selection method for MCI classification in this work. It unifies the above measures into a minimization problem associated with a least square error and an Elastic Net regularizer. Experimental results demonstrate that the diagnosis accuracy for MCI subjects can be significantly improved using our proposed feature selection method.

Cervical cancer is one of the primary factors that endanger women’s health, and Thinprep cytologic test (TCT) is the common testing tool for the early diagnosis of cervical cancer. However, it is tedious and time-consuming for pathologists to assess and find abnormal cells in many TCT samples. Thus, automatic detection of abnormal cervical cells is highly demanded. Nevertheless, false positive cells are inevitable after automatic detection. It is still a burden for the pathologist if the false positive rate is high. To this end, here we propose a semi-supervised cervical cell diagnosis method that can significantly reduce the false positive rate. First, we incorporate a detection network to localize the suspicious abnormal cervical cells. Then, we design a semi-supervised classification network to identify whether the cervical cells are truly abnormal or not. To boost the performance of the semi-supervised classification network, and make full use of the localizing information derived from the detection network, we use the predicted bounding boxes of the detection network as an additional constraint for the attention masks from the classification network. Besides, we also develop a novel consistency constraint between the teacher and student models to guarantee the robustness of the network. Our experimental results show that our network can achieve satisfactory classification accuracy using only a limited number of labeled cells, and also greatly reduce the false positive rate in cervical cell detection.

Stratigraphic correlation is essential in field evaluation as it provides the necessary tops to compartmentalize the reservoir. It further contributes to other parts of the field development planning cycle such as reservoir modeling, volumetric assessment, production allocation, etc. Traditional approach of manual pairwise correlation is labor-intensive and time-consuming. This research presents a novel automated stratigraphic correlator to create well top and zonation interpretations using supervised machine learning algorithms of Convolutional- and Recurrent-Neural-Networks (CNNs and RNNs). An automated stratigraphic correlator is created that enables stratigraphic well top and zonation interpretations learned from the well logs of a subset of wells with zonation information manually provided by human experts. The method can efficiently learn the patterns and hidden information from the well logs’ sequential data, implicitly capture the domain expertise, and streamline and automate the traditional manual repetitive work. Our method supersedes existing approaches like Multiple Sequence Alignment (MSA) by incorporating domain expertise through tops/zones picked by geologists. A Bidirectional Long Short-Term Memory (BiLSTM) is used to interpret the log data, since deposition by nature is a sequential process and RNNs can intrinsically capture such series. An Inception autoencoder CNN is also applied in this workflow for stratigraphic interpretation. Reliable post-processing is also included using the predicted zone probability logs to quantify the overall confidence score of well zonation, and to correct misinterpretation when necessary using transition frequencies in log data through a linear chain graphical probabilistic model. The methodology is tested on one of the major Middle East oilfields with around 1,500 wells to prove its efficiency and capability. The overall methodology involves data pre-processing, deep learning model training and prediction, and the post-processing of model-predicted results. In this specific workflow, the machine learning targets include both the prediction of zones (multi-class classification/segmentation problem) and the prediction of well tops (edge-detection problem). Thus, a supervised multi-task learning on a single field using CNNs and RNNs is implemented to be able to perform different tasks with the same model. The inputs to the training module include trajectory logs and other measured logs such as gamma-ray, resistivity, neutron density, etc. All inputs are normalized to zero mean and unit standard deviation. For wells with missing log values, the approach can either discard it or perform data imputation to reconstruct the data using different automated algorithms. The machine learning engine uses two different algorithms (BiLSTM and Inception autoencoder CNN), with many other deep learning models tested. The training loss function includes zone categorical cross entropy loss, tops edge detection binary cross entropy loss and L2-norm regularization term. The learning rate is dynamically adjusted during training so that it is reduced when the loss is stalled. The post-processing uses the machine learning predicted zone probability logs to select the zoning sequence that maximizes overall zonation probability and treats it as the confidence score of well zonation. This dramatically helps in constraining the outcome stratigraphic interpretation by geological succession and minimizing the correlation error. The entire workflow has been applied to one major Middle East oilfield with a large number of pre-interpreted well logs, with 60% of the wells used to train the deep learning models, 20% used for validation and the rest are for blind test. Both BiLSTM and Inception autoencoder CNN show close to human-level performance in the blind test dataset. The mean absolute error of well tops interpretation after post-processing is around 3 m throughout all analyzed wells, which provided an accuracy of nearly 90% for the blind test dataset. The classification precision and accuracy also demonstrate close-to-human-level performance in the major zones with sufficient data. It has been noticed that for cases without missing data, Inception autoencoder CNN achieves best performance, while BiLSTM benefits a lot from imputation when missing data exists. The methodology automates and streamlines the originally time-consuming stratigraphic correlation process. It performs better than existing approaches through a well-developed machine learning framework with comprehensive data pre- and post-processing. The resulting stratigraphic correlation proves to be extremely reliable even with a small number of seed wells, and it requires minimal user intervention during the process. Through deep learning techniques such as transfer learning, the proposed methodology can be readily applied to other fields even with limited training data.

High-quality magnetic resonance (MR) image, i.e., with near isotropic voxel spacing, is desirable in various scenarios of medical image analysis. However, many MR images are acquired using good in-plane resolution but large spacing between slices in clinical practice. In this work, we propose a novel deep-learning-based super-resolution algorithm to generate high-resolution (HR) MR images of small slice spacing from low-resolution (LR) inputs of large slice spacing. Notice that real HR images are needed in most existing deep-learning-based methods to supervise the training, but in clinical scenarios, usually they will not be acquired. Therefore, our unique goal herein is to design and train the super-resolution network without real HR ground-truth. Specifically, two-staged training is used in our method. In the first stage, HR images of reduced slice spacing are synthesized from real LR images using variational auto-encoder (VAE). Although these synthesized HR images of reduced slice spacing are as realistic as possible, they may still suffer from unexpected morphing induced by VAE, implying that the synthesized HR images cannot be paired with the real LR images in terms of anatomical structure details. In the second stage, we degrade the synthesized HR images to generate corresponding LR-HR image pairs and train a super-resolution network based on these synthesized pairs. The underlying mechanism is that such a super-resolution network is less vulnerable to anatomical variability. Experiments on knee MR images successfully demonstrate the effectiveness of our proposed solution to reduce the slice spacing for better rendering.

Biomedical imaging Is an important source of information in cancer research. Characterizations of cancer morphology at onset, progression, and in response to treatment provide complementary information to that gleaned from genomics and clinical data. Accurate extraction and classification of both visual and latent image features Is an increasingly complex challenge due to the increased complexity and resolution of biomedical image data. In this paper, we present four deep learning-based image analysis methods from the Computational Precision Medicine (CPM) satellite event of the 21st International Medical Image Computing and Computer Assisted Intervention (MICCAI 2018) conference. One method Is a segmentation method designed to segment nuclei in whole slide tissue images (WSIs) of adult diffuse glioma cases. It achieved a Dice similarity coefficient of 0.868 with the CPM challenge datasets. Three methods are classification methods developed to categorize adult diffuse glioma cases into oligodendroglioma and astrocytoma classes using radiographic and histologic image data. These methods achieved accuracy values of 0.75, 0.80, and 0.90, measured as the ratio of the number of correct classifications to the number of total cases, with the challenge datasets. The evaluations of the four methods indicate that (1) carefully constructed deep learning algorithms are able to produce high accuracy in the analysis of biomedical image data and (2) the combination of radiographic with histologic image information improves classification performance.

Brain magnetic resonance (MR) segmentation for hydrocephalus patients is considered as a challenging work. Encoding the variation of the brain anatomical structures from different individuals cannot be easily achieved. The task becomes even more difficult especially when the image data from hydrocephalus patients are considered, which often have large deformations and differ significantly from the normal subjects. Here, we propose a novel strategy with hard and soft attention modules to solve the segmentation problems for hydrocephalus MR images. Our main contributions are three-fold: 1) the hard-attention module generates coarse segmentation map using multi-atlas-based method and the VoxelMorph tool, which guides subsequent segmentation process and improves its robustness; 2) the soft-attention module incorporates position attention to capture precise context information, which further improves the segmentation accuracy; 3) we validate our method by segmenting insula, thalamus and many other regions-of-interests (ROIs) that are critical to quantify brain MR images of hydrocephalus patients in real clinical scenario. The proposed method achieves much improved robustness and accuracy when segmenting all 17 consciousness-related ROIs with high variations for different subjects. To the best of our knowledge, this is the first work to employ deep learning for solving the brain segmentation problems of hydrocephalus patients., ISBI 2020

Vessel segmentation and anatomical labeling are of great significance for vascular disease analysis. Because vessels in 3D images are the tree-like tubular structures with diverse shapes and sizes, and direct use of convolutional neural networks (CNNs, based on spatial convolutional kernels) for vessel segmentation often encounters great challenges. To tackle this problem, we propose a graph convolutional network (GCN)-based point cloud approach to improve vessel segmentation over the conventional CNN-based method and further conduct semantic labeling on 13 major head and neck vessels. The proposed method can not only learn the global shape representation but also precisely adapt to local vascular shapes by utilizing the prior knowledge of tubular structures to explicitly learn anatomical shape. Specifically, starting from rough segmentation using V-Net, our approach further refines the segmentation and performs labeling on the refined segmentations, with two steps. First, a point cloud network is applied to the points formed by initial vessel voxels to refine vessel segmentation. Then, GCN is employed on the point cloud to further label vessels into 13 major segments. To evaluate the performance of our proposed method, CT angiography images (covering heads and necks) of 72 subjects are used in our experiment. Using four-fold cross-validation, an average Dice coefficient of 0.965 can be achieved for vessel segmentation compared to that of 0.885 obtained by the conventional V-Net based segmentation. Also, for vessel labeling, our proposed algorithm achieves an average Dice coefficient of 0.899 for 13 vessel segments compared to that of 0.829 by V-Net. These results show that our proposed method could facilitate head and neck vessel analysis by providing automatic and accurate vessel segmentation and labeling results.

MR images of knee joint are usually collected in axial, coronal, and sagittal views with large slice spacing for clinical study. Current methods either segment images in different views separately or apply super-resolution fusion before 3D segmentation. Knee images segmentation transfer between different views is still an open problem. Moreover, the majority of manual labelling works focus on the sagittal-view, and practically it is hard to collect label maps for the coronal- and axial-views, which are also invaluable for observing knee injuries. In this paper, we propose a novel algorithm to transfer sagittal-view annotations to the other views. First, we build a supervised low-resolution segmentation (LR-Seg) module based on the down-sampled sagittal-view slices to obtain the label map on the target view. And then a context transfer module is proposed to refine the segmentations using target-view context. Then by iterative learning of these two modules, the context from one result can be used to guide the training of the other. Experimental results show that our algorithm can greatly alleviate the burden of manually labeling works from clinicians and gain comparable segmentation results on axial and coronal views.

Computed Tomography Perfusion (CTP) images have drawn extensive attention in acute ischemic stroke assessment due to its imaging speed and ability to provide dynamic perfusion quantification. However, the cerebral ischemic infarcted core has high individual variability and low contrast, and multiple CTP parametric maps need to be referred for precise delineation of the core region. It has thus become a challenging task to develop automatic segmentation algorithms. The widely applied segmentation algorithms such as U-Net lack specific modeling for image subtype in the dataset, and thus the performance remains unsatisfactory. In this paper, we propose a novel cluster-representation learning approach to address these difficulties. Specifically, we first cluster the training samples based on their similarities of the segmentation difficulty. Each cluster represents a different subtype of training images and is then used to train its own cluster-representative model. The models will be capable of extracting cluster-representative features from training samples as clustering priors, which are further fused into an overall segmentation model (for all training samples). The fusion mechanism is able to adaptively select optimal subset(s) of clustering priors which can further guide the segmentation of each unseen testing image and reduce influences from high variability of CTP images. We have applied our method on 94 subjects of ISLES 2018 dataset. By comparing with the baseline U-Net, the experiments have shown an absolute increase of 8% in Dice score and a reduction of 10mm in Hausdorff Distance for ischemic infarcted core segmentation. This method can also be generalized to other U-Net-like architectures to further improve their representative capacity.

Infection segmentation is essential for quantitative assessment in computer-aided management of COVID-19. However, clinical CT images are usually heterogeneous, which are reconstructed by different protocols with varying appearance and voxel spacing due to radiologist preference. Most existing infection segmentation models are only trained using specific types of CT images, which would undermine the performance when applied to other types of CT images. Therefore, it is highly desirable to construct a model that could be applied to heterogeneous CT images in the urgent COVID-19 applications. In this paper, we present a two-stage mapping-segmentation framework for delineating COVID-19 infections from CT images. To compensate for heterogeneity of CT images obtained from different imaging centers, we develop an image-level domain-adaptive process to transform all kinds of images into a target type, and then segment COVID-19 infections accordingly. Experiments show that the infection delineation performance based on our proposed method is superior to the model trained jointly using mixture of all types of images, and is also comparable to those models supervised by using only specific types of images (if applicable).

Radiomics is widely-used in imaging based clinical studies as a way of extracting high-throughput image descriptors. However, current tools for extracting radiomics features are generally run on CPU only, which leads to large time consumption in situations such as large datasets or complicated task/method verifications. To address this limitation, we have developed a GPU based toolkit namely cuRadiomics, where the computing time can be significantly reduced. In cuRadiomics, the CUDA-based feature extraction process for two different classes of radiomics features, including 18 first-order features based on intensity histograms and 23 texture features based on gray level cooccurrence matrix (GLCM), has been developed. We have demonstrated the advantage of the cuRadiomics toolkit over CPU-based feature extraction methods using BraTS18 and KiTS19 datasets. For example, regarding the whole image as ROI, feature extraction process using cuRadiomics is 143.13 times faster than that using PyRadiomics. Thus, the potential advantage provided by cuRadiomics enables the radiomics related statistical methods more adaptive and convenient to use than before. Our proposed cuRadiomics toolkit is now publicly available at https://github.com/jiaoyining/cuRadiomics.

Knee osteoarthritis (OA) is one of the most common musculoskeletal disorders and requires early-stage diagnosis. Nowadays, the deep convolutional neural networks have achieved greatly in the computer-aided diagnosis field. However, the construction of the deep learning models usually requires great amounts of annotated data, which is generally high-cost. In this paper, we propose a novel approach for knee cartilage defects assessment, including severity classification and lesion localization. This can be treated as a subtask of knee OA diagnosis. Particularly, we design a self-ensembling framework, which is composed of a student network and a teacher network with the same structure. The student network learns from both labeled data and unlabeled data and the teacher network averages the student model weights through the training course. A novel attention loss function is developed to obtain accurate attention masks. With dual-consistency checking of the attention in the lesion classification and localization, the two networks can gradually optimize the attention distribution and improve the performance of each other, whereas the training relies on partially labeled data only and follows the semi-supervised manner. Experiments show that the proposed method can significantly improve the self-ensembling performance in both knee cartilage defects classification and localization, and also greatly reduce the needs of annotated data.

Sparse-data problem is a main issue that influences the performances of statistical language models; statistical language model based on word classes is an effective method to solve sparse-data problems. This paper presents a definition of word similarity by utilizing mutual information of adjoining words, and gives the definition of word set similarity based on word similarity, and puts forward a bottom-up hierarchical word clustering algorithm which can get global optimum. Experimental results show that the word clustering algorithm is of high executing speed and have good clustering performances. We then interpolated the class-based models with the word-based models and found that it mitigates remaining sparse-data problems of statistical language models.

Population studies of brain function with resting-state functional magnetic resonance imaging (rs-fMRI) rely on accurate intersubject registration of functional areas. This is typically achieved through registration using high-resolution structural images with more spatial details and better tissue contrast. However, accumulating evidence has suggested that such strategy cannot align functional regions well because functional areas are not necessarily consistent with anatomical structures. To alleviate this problem, a number of registration algorithms based directly on rs-fMRI data have been developed, most of which utilize functional connectivity (FC) features for registration. However, most of these methods usually extract functional features only from the thin and highly curved cortical grey matter (GM), posing great challenges to accurate estimation of whole-brain deformation fields. In this article, we demonstrate that additional useful functional features can also be extracted from the whole brain, not restricted to the GM, particularly the white-matter (WM), for improving the overall functional registration. Specifically, we quantify local anisotropic correlation patterns of the blood oxygenation level-dependent (BOLD) signals using tissue-specific patch-based functional correlation tensors (ts-PFCTs) in both GM and WM. Functional registration is then performed by integrating the features from different tissues using the multi-channel large deformation diffeomorphic metric mapping (mLDDMM) algorithm. Experimental results show that our method achieves superior functional registration performance, compared with conventional registration methods.

Category-based statistic language model is an important method to solve the problem of sparse data in statistical language models. But there are two bottlenecks about this model: 1) The problem of word clustering, it is hard to find a suitable clustering method that has good performance and has not large amount of computation; 2) Class-based method always loses some prediction ability to adapt the text of different domain. In order to solve above problems, a novel definition of word similarity by utilizing mutual information was presented. Based on word similarity, the definition of word set similarity was given and a bottom-up hierarchical clustering algorithm was proposed. Experimental results show that the word clustering algorithm based on word similarity is better than conventional greedy clustering method in speed and performance, the perplexity is reduced from 283 to 207.8.

Statistic language model based on classes of words is an effective method to solve sparse-data problems. Conventional statistical word clustering models usually base on greedy principle. The common Metric for evaluating a clustering model is the perplexity of the corpus or likelihood function. Conventional statistical word clustering algorithms often converge to a local optimum, so global optimum is not guaranteed, and initial choices can influence the final clustering result. Pointing to these problems, this paper presents a definition of word similarity by utilizing mutual information, and gives the definition of word set similarity based on word similarity, and puts forward a bottom-up hierarchical word clustering algorithm which can get global optimum. Experimental results show that the word clustering algorithm is of high executing speed and has good clustering performance. We then interpolated the class-based models with the word-based models and found that it mitigates the remaining sparse-data problems.

Medical image segmentation is challenging especially in dealing with small dataset of 3D MR images. Encoding the variation of brain anatomical structures from individual subjects cannot be easily achieved, which is further challenged by only a limited number of well labeled subjects for training. In this study, we aim to address the issue of brain MR image segmentation in small dataset. First, concerning the limited number of training images, we adopt adversarial defense to augment the training data and therefore increase the robustness of the network. Second, inspired by the prior knowledge of neural anatomies, we reorganize the segmentation tasks of different regions into several groups in a hierarchical way. Third, the task reorganization extends to the semantic level, as we incorporate an additional object-level classification task to contribute high-order visual features toward the pixel-level segmentation task. In experiments we validate our method by segmenting gray matter, white matter, and several major regions on a challenge dataset. The proposed method with only seven subjects for training can achieve 84.46% of Dice score in the onsite test set.

Semantic segmentation is essentially important to biomedical image analysis. Many recent works mainly focus on integrating the Fully Convolutional Network (FCN) architecture with sophisticated convolution implementation and deep supervision. In this paper, we propose to decompose the single segmentation task into three subsequent sub-tasks, including (1) pixel-wise image segmentation, (2) prediction of the class labels of the objects within the image, and (3) classification of the scene the image belonging to. While these three sub-tasks are trained to optimize their individual loss functions of different perceptual levels, we propose to let them interact by the task-task context ensemble. Moreover, we propose a novel sync-regularization to penalize the deviation between the outputs of the pixel-wise segmentation and the class prediction tasks. These effective regularizations help FCN utilize context information comprehensively and attain accurate semantic segmentation, even though the number of the images for training may be limited in many biomedical applications. We have successfully applied our framework to three diverse 2D/3D medical image datasets, including Robotic Scene Segmentation Challenge 18 (ROBOT18), Brain Tumor Segmentation Challenge 18 (BRATS18), and Retinal Fundus Glaucoma Challenge (REFUGE18). We have achieved top-tier performance in all three challenges., Comment: IEEE Transactions on Medical Imaging

Automatic dense parcellation of brain MR image, which labels hundreds of regions of interest (ROIs), plays an important role for neuroimage analysis. Specifically, the brain image parcellation using deep learning technology has been widely recognized for its effective performance, but it remains limited in actual application due to its high demand for sufficient training data and intensive memory allocation of GPU resources. Due to the high cost of manual segmentation, it is usually not feasible to provide large dataset for training the network. On the other hand, it is relatively easy to transfer labeling information to many new unlabeled datasets and thus augment the training data. However, the augmented data can only be considered as weakly labeled for training. Therefore, in this paper, we propose a cascaded weakly super- vised confidence integration network (CINet). The main contributions of our method are two-folds. First, we propose the image registration-based data argumentation method, and evaluate the confidence of the labeling information for each augmented image. The augmented data, as well as the original yet small training dataset, contribute to the modeling of the CINet jointly for segmentation. Second, we propose the random crop strategy to handle the large amount of feature channels in the network, which are needed to label hundreds of neural ROIs. The demanding requirement to GPU memory is thus relieved, while better accuracy can also be achieved. In experiments, we use 37 manually labeled subjects and augment 96 images with weak labels for training. The testing result in overall Dice score over 112 brain regions reaches 75%, which is higher than using the original training data only.

Limb salvage surgery of malignant pelvic tumors is the most challenging procedure in musculoskeletal oncology due to the complex anatomy of the pelvic bones and soft tissues. It is crucial to accurately resect the pelvic tumors with appropriate margins in this procedure. However, there is still a lack of efficient and repetitive image planning methods for tumor identification and segmentation in many hospitals. In this paper, we present a novel deep learning-based method to accurately segment pelvic bone tumors in MRI. Our method uses a multi-view fusion network to extract pseudo-3D information from two scans in different directions and improves the feature representation by learning a relational context. In this way, it can fully utilize spatial information in thick MRI scans and reduce over-fitting when learning from a small dataset. Our proposed method was evaluated on two independent datasets collected from 90 and 15 patients, respectively. The segmentation accuracy of our method was superior to several comparing methods and comparable to the expert annotation, while the average time consumed decreased about 100 times from 1820.3 seconds to 19.2 seconds. In addition, we incorporate our method into an efficient workflow to improve the surgical planning process. Our workflow took only 15 minutes to complete surgical planning in a phantom study, which is a dramatic acceleration compared with the 2-day time span in a traditional workflow.

Deformable brain MR image registration is challenging due to large inter-subject anatomical variation. For example, the highly complex cortical folding pattern makes it hard to accurately align corresponding cortical structures of individual images. In this paper, we propose a novel deep learning way to simplify the difficult registration problem of brain MR images. Specifically, we train a morphological simplification network (MS-Net), which can generate a simple image with less anatomical details based on the complex input. With MS-Net, the complexity of the fixed image or the moving image under registration can be reduced gradually, thus building an individual (simplification) trajectory represented by MS-Net outputs. Since the generated images at the ends of the two trajectories (of the fixed and moving images) are so simple and very similar in appearance, they are easy to register. Thus, the two trajectories can act as a bridge to link the fixed and the moving images, and guide their registration. Our experiments show that the proposed method can achieve highly accurate registration performance on different datasets (i.e., NIREP, LPBA, IBSR, CUMC, and MGH). Moreover, the method can be also easily transferred across diverse image datasets and obtain superior accuracy on surface alignment. We propose MS-Net as a powerful and flexible tool to simplify brain MR images and their registration. To our knowledge, this is the first work to simplify brain MR image registration by deep learning, instead of estimating deformation field directly.

Parkinson's disease (PD) is a neurodegenerative disorder that progressively hampers the brain functions and leads to various movement and non-motor symptoms. However, it is difficult to attain early-stage PD diagnosis based on the subjective judgment of physicians in clinical routines. Therefore, automatic and accurate diagnosis of PD is highly demanded, so that the corresponding treatment can be implemented more appropriately. In this paper, we focus on finding the most discriminative features from different brain regions in PD through T1-weighted MR images, which can help the subsequent PD diagnosis. Specifically, we proposed a novel iterative canonical correlation analysis (ICCA) feature selection method, aiming at exploiting MR images in a more comprehensive manner and fusing features of different types into a common space. To state succinctly, we first extract the feature vectors from the gray matter and the white matter tissues separately, represented as insights of two different anatomical feature spaces for the subject's brain. The ICCA feature selection method aims at iteratively finding the optimal feature subset from two sets of features that have inherent high correlation with each other. In experiments we have conducted thorough investigations on the optimal feature set extracted by our ICCA method. We also demonstrate that using the proposed feature selection method, the PD diagnosis performance is further improved, and also outperforms many state-of-the-art methods.

Brain tumor grading is pivotal in treatment planning. Contrast-enhanced T1-weighted MR image is commonly used for grading. However, the classification of different types of high-grade gliomas using T1-weighted MR images is still challenging, due to the lack of imaging biomarkers. Previous studies only focused on simple visual features, ignoring rich information provided by MR images. In this paper, we propose an automatic classification pipeline using random forest to differentiate the WHO Grade III and Grade IV gliomas, by extracting discriminative features based on 3D patches. The proposed pipeline consists of three main steps in both the training and the testing stages. First, we select numerous 3D patches in and around the tumor regions of the given MR images. This can suppress the intensity information from the normal region, which is trivial for the classification process. Second, we extract features based on both patch-wise information and subject-wise clinical information, and then we refine this step to optimize the performance of malignant tumor classification. Third, we incorporate the classification forest for training/testing the classifier. We validate the proposed framework on 96 malignant brain tumor patients that consist of both Grade III (N = 38) and Grade IV gliomas (N = 58). The experiments show that the proposed framework has demonstrated its validity in the application of high-grade gliomas classification, which may help improve the poor prognosis of high-grade gliomas.

It is fundamentally important to fuse the brain atlas from magnetic resonance (MR) images for many imaging-based studies. Most existing works focus on fusing the atlases from high-quality MR images. However, for low-quality diagnostic images (i.e., with high inter-slice thickness), the problem of atlas fusion has not been addressed yet. In this paper, we intend to fuse the brain atlas from the high-thickness diagnostic MR images that are prevalent for clinical routines. The main idea of our works is to extend the conventional groupwise registration by incorporating a novel super-resolution strategy. The contribution of the proposed super-resolution framework is two-fold. First, each high-thickness subject image is reconstructed to be isotropic by the patch-based sparsity learning. Then, the reconstructed isotropic image is enhanced for better quality through the random-forest-based regression model. In this way, the images obtained by the super-resolution strategy can be fused together by applying the groupwise registration method to construct the required atlas. Our experiments have shown that the proposed framework can effectively solve the problem of atlas fusion from the low-quality brain MR images.

Recently, more and more attention is drawn to the field of medical image synthesis across modalities. Among them, the synthesis of computed tomography (CT) image from T1-weighted magnetic resonance (MR) image is of great importance, although the mapping between them is highly complex due to large gaps of appearances of the two modalities. In this work, we aim to tackle this MR-to-CT synthesis task by a novel deep embedding convolutional neural network (DECNN). Specifically, we generate the feature maps from MR images, and then transform these feature maps forward through convolutional layers in the network. We can further compute a tentative CT synthesis from the midway of the flow of feature maps, and then embed this tentative CT synthesis result back to the feature maps. This embedding operation results in better feature maps, which are further transformed forward in DECNN. After repeating this embedding procedure for several times in the network, we can eventually synthesize a final CT image in the end of the DECNN. We have validated our proposed method on both brain and prostate imaging datasets, by also comparing with the state-of-the-art methods. Experimental results suggest that our DECNN (with repeated embedding operations) demonstrates its superior performances, in terms of both the perceptive quality of the synthesized CT image and the run-time cost for synthesizing a CT image.

Functional magnetic resonance imaging (fMRI) measures changes in blood-oxygenation-level-dependent (BOLD) signals to detect brain activities. It has been recently reported that the spatial correlation patterns of resting-state BOLD signals in the white matter (WM) also give WM information often measured by diffusion tensor imaging (DTI). These correlation patterns can be captured using functional correlation tensor (FCT), which is analogous to the diffusion tensor (DT) obtained from DTI. In this paper, we propose a noise-robust FCT method aiming at further improving its quality, and making it eligible for further neuroscience study. The novel FCT estimation method consists of three major steps: First, we estimate the initial FCT using a patch-based approach for BOLD signal correlation to improve the noise robustness. Second, by utilizing the relationship between functional and diffusion data, we employ a regression forest model to learn the mapping between the initial FCTs and the corresponding DTs using the training data. The learned forest can then be applied to predict the DTI-like tensors given the initial FCTs from the testing fMRI data. Third, we re-estimate the enhanced FCT by utilizing the DTI-like tensors as a feedback guidance to further improve FCT computation. We have demonstrated the utility of our enhanced FCTs in Alzheimer's disease (AD) diagnosis by identifying mild cognitive impairment (MCI) patients from normal subjects.

Population studies of brain function with resting-state functional magnetic resonance imaging (rs-fMRI) largely rely on the accurate inter-subject registration of functional areas. This is typically achieved through registration of the corresponding T1-weighted MR images with more structural details. However, accumulating evidence has suggested that such strategy cannot well-align functional regions which are not necessarily confined by the anatomical boundaries defined by the T1-weighted MR images. To mitigate this problem, various registration algorithms based directly on rs-fMRI data have been developed, most of which have utilized functional connectivity (FC) as features for registration. However, most of the FC-based registration methods usually extract the functional features only from the thin and highly curved cortical grey matter (GM), posing a great challenge in accurately estimating the whole-brain deformation field. In this paper, we demonstrate that the additional useful functional features can be extracted from brain regions beyond the GM, particularly, white-matter (WM) based on rs-fMRI, for improving the overall functional registration. Specifically, we quantify the local anisotropic correlation patterns of the blood oxygenation level-dependent (BOLD) signals, modeled by functional correlation tensors (FCTs), in both GM and WM. Functional registration is then performed based on multiple components of the whole-brain FCTs using a multichannel Large Deformation Diffeomorphic Metric Mapping (mLDDMM) algorithm. Experimental results show that our proposed method achieves superior functional registration performance, compared with other conventional registration methods.

Automatic labeling of anatomical structures in brain images plays an important role in neuroimaging analysis. Among all methods, multi-atlas based segmentation methods are widely used, due to their robustness in propagating prior label information. However, non-linear registration is always needed, which is time-consuming. Alternatively, the patch-based methods have been proposed to relax the requirement of image registration, but the labeling is often determined independently by the target image information, without getting direct assistance from the atlases. To address these limitations, in this paper, we propose a multi-atlas guided 3D fully convolutional networks (FCN) for brain image labeling. Specifically, multi-atlas based guidance is incorporated during the network learning. Based on this, the discriminative of the FCN is boosted, which eventually contribute to accurate prediction. Experiments show that the use of multi-atlas guidance improves the brain labeling performance.

It has been recently demonstrated that the local BOLD signals in resting-state fMRI (rs-fMRI) can be captured for the white matter (WM) by functional correlation tensors (FCTs). FCTs provide similar orientation information as diffusion tensors (DTs), and also functional information concerning brain dynamics. However, FCTs are susceptible to noise due to the low signal-to-noise ratio nature of WM BOLD signals. Here we introduce a robust FCT estimation method to facilitate individualized diagnosis. First, we develop a noise-tolerating patch-based approach to measure spatiotemporal correlations of local BOLD signals. Second, it is also enhanced by DTs predicted from the input rs-fMRI using a learning-based regression model. We evaluate our trained regressor using the high-resolution HCP dataset. The regressor is then applied to estimate the robust FCTs for subjects in the ADNI2 dataset. We demonstrate for the first time the disease diagnostic value of robust FCTs.

In this paper we show how to use blind source separation to estimate shape from polarised images from a fixed viewing and illumination direction. Our method does not require prior knowledge of the polariser angles nor the refractive index. We use weighted singular vector decomposition (SVD) to estimate the unknown parameters, where the weights are dictated by the physics of polarisation. In a two-step process, we firstly use least square fitting to obtain the polariser angles, and then estimate the refractive index using a mutual information criterion function. We show that the proposed method is capable of calculating robust shape information compared with alternative approaches relying on the same input information. The proposed method can be applied when using uncalibrated polarisation filters. It can also be used effectively when the subject moves during image capture.

Automatic labeling of the hippocampus in brain MR images is highly demanded, as it has played an important role in imaging-based brain studies. However, accurate labeling of the hippocampus is still challenging, partially due to the ambiguous intensity boundary between the hippocampus and surrounding anatomies. In this paper, we propose a concatenated set of spatially-localized random forests for multi-atlas-based hippocampus labeling of adult/infant brain MR images. The contribution in our work is two-fold. First, each forest classifier is trained to label just a specific sub-region of the hippocampus, thus enhancing the labeling accuracy. Second, a novel forest selection strategy is proposed, such that each voxel in the test image can automatically select a set of optimal forests, and then dynamically fuses their respective outputs for determining the final label. Furthermore, we enhance the spatially-localized random forests with the aid of the auto-context strategy. In this way, our proposed learning framework can gradually refine the tentative labeling result for better performance. Experiments show that, regarding the large datasets of both adult and infant brain MR images, our method owns satisfactory scalability by segmenting the hippocampus accurately and efficiently.

Purpose: Automatic brainimage labeling is highly demanded in the field of medical imageanalysis. Multiatlas-based approaches are widely used due to their simplicity and robustness in applications. Also, random forest technique is recognized as an efficient method for labeling, although there are several existing limitations. In this paper, the authors intend to address those limitations by proposing a novel framework based on the hierarchical learning of atlas forests. Methods: Their proposed framework aims to train a hierarchy of forests to better correlate voxels in the MR images with their corresponding labels. There are two specific novel strategies for improving brainimage labeling. First, different from the conventional ways of using a single level of random forests for brain labeling, the authors design a hierarchical structure to incorporate multiple levels of forests. In particular, each atlas forest in the bottom level is trained in accordance with each individual atlas, and then the bottom-level forests are clustered based on their capabilities in labeling. For each clustered group, the authors retrain a new representative forest in the higher level by using all atlases associated with the lower-level atlas forests in the current group, as well as the tentative label maps yielded from the lower level. This clustering and retraining procedure is conducted iteratively to yield a hierarchical structure of forests. Second, in the testing stage, the authors also present a novel atlas forest selection method to determine an optimal set of atlas forests from the constructed hierarchical structure (by disabling those nonoptimal forests) for accurately labeling the test image. Results: For validating their proposed framework, the authors evaluate it on the public datasets, including Alzheimer’s disease neuroimaging initiative, Internet brain segmentation repository, and LONI LPBA40. The authors compare the results with the conventional approaches. The experiments show that the use of the two novel strategies can significantly improve the labeling performance. Note that when more levels are constructed in the hierarchy, the labeling performance can be further improved, but more computational time will be also required. Conclusions: The authors have proposed a novel multiatlas-based framework for automatic and accurate labeling of brain anatomies, which can achieve accurate labeling results for MR brainimages.

Parkinson’s disease (PD) is a major progressive neurodegenerative disorder. Accurate diagnosis of PD is crucial to control the symptoms appropriately. However, its clinical diagnosis mostly relies on the subjective judgment of physicians and the clinical symptoms that often appear late. Recent neuroimaging techniques, along with machine learning methods, provide alternative solutions for PD screening. In this paper, we propose a novel feature selection technique, based on iterative canonical correlation analysis (ICCA), to investigate the roles of different brain regions in PD through T1-weighted MR images. First of all, gray matter and white matter tissue volumes in brain regions of interest are extracted as two feature vectors. Then, a small group of significant features were selected using the iterative structure of our proposed ICCA framework from both feature vectors. Finally, the selected features are used to build a robust classifier for automatic diagnosis of PD. Experimental results show that the proposed feature selection method results in better diagnosis accuracy, compared to the baseline and state-of-the-art methods.

Category-based statistical language model is an important method to solve the problem of sparse data, but there are two bottlenecks about this model: (1) the problem of word clustering, it is hard to find a suitable clustering method that has good performance and has not large amount of computation. (2) class-based method always loses some prediction ability to adapt the text of different domain. In order to solve above problems, a definition of word similarity by utilizing mutual information is presented. Based on word similarity, the definition of word set similarity is given. Experiments show that word clustering algorithm based on similarity is better than conventional greedy clustering method in speed and performance, the perplexity is reduced from 283 to 218.

It is important, yet a challenging procedure, to segment bones and cartilages from knee MR images. In this paper, we propose multi-atlas context forests to first segment bones and then segment cartilages. Specifically, for both the bone and cartilage segmentations, we iteratively train sets of random forests, based on training atlas images, to classify the individual voxels. The random forests rely on 1 the appearance features directly computed from images and also 2 the context features associated with tentative segmentation results, generated by the previous layer of random forest in the iterative framework. To extract context features, multiple atlases with expert segmentation are first registered, with the tentative segmentation result of the subject under consideration. Then, the spatial priors of anatomical labels of registered atlases are computed and used to calculate context features of the subject. Note that these multi-atlas context features will be iteratively refined based on the updated tentative segmentation result of the subject. As better segmentation result leads to more accurate registration between multiple atlases and the subject, context features will become increasingly more useful for the training of subsequent random forests in the iterative framework. As validated by experiments on the SKI10 dataset, our proposed method can achieve high segmentation accuracy.