Retinal Vessel Segmentation by A Transformer-U-Net Hybrid Model with Dual-Path Decoder.

This paper introduces an effective and efficient framework for retinal vessel segmentation. First, we design a Transformer-CNN hybrid model in which a Transformer module is inserted inside the U-Net to capture long-range interactions. Second, we design a dual-path decoder in the U-Net framework, which contains two decoding paths for multi-task outputs. Specifically, we train the extra decoder to predict vessel skeletons as an auxiliary task which helps the model learn balanced features. The proposed framework, named as TSNet, not only achieves good performances in a fully supervised learning manner but also enables a rough skeleton annotation process. The annotators only need to roughly delineate vessel skeletons instead of giving precise pixel-wise vessel annotations. To learn with rough skeleton annotations plus a few precise vessel annotations, we propose a skeleton semi-supervised learning scheme. We adopt a mean teacher model to produce pseudo vessel annotations and conduct annotation correction for roughly labeled skeletons annotations. This learning scheme can achieve promising performance with fewer annotation efforts. We have evaluated TSNet through extensive experiments on five benchmarking datasets. Experimental results show that TSNet yields state-of-the-art performances on retinal vessel segmentation and provides an efficient training scheme in practice.

A Dual Coordinate System Vertebra Landmark Detection Network with Sparse-to-Dense Vertebral Line Interpolation.

Precise surveillance and assessment of spinal disorders are important for improving health care and patient survival rates. The assessment of spinal disorders, such as scoliosis assessment, depends heavily on precise vertebra landmark localization. However, existing methods usually search for only a handful of keypoints in a high-resolution image. In this paper, we propose the S2D-VLI VLDet network, a unified end-to-end vertebra landmark detection network for the assessment of scoliosis. The proposed network considers the spatially relevant information both from inside and between vertebrae. The new vertebral line interpolation method converts the training labels from sparse to dense, which can improve the network learning process and method performance. In addition, through the combined use of the Cartesian and polar coordinate systems in our method, the symmetric mean absolute percentage error (SMAPE) in scoliosis assessment can be reduced substantially. Specifically, as shown in the experiments, the SMAPE value decreases from 9.82 to 8.28. The experimental results indicate that our proposed approach is beneficial for estimating the Cobb angle and identifying landmarks in X-ray scans with low contrast.

Unsupervised Domain Adaptation for Medical Image Segmentation Using Transformer With Meta Attention.

Image segmentation is essential to medical image analysis as it provides the labeled regions of interest for the subsequent diagnosis and treatment. However, fully-supervised segmentation methods require high-quality annotations produced by experts, which is laborious and expensive. In addition, when performing segmentation on another unlabeled image modality, the segmentation performance will be adversely affected due to the domain shift. Unsupervised domain adaptation (UDA) is an effective way to tackle these problems, but the performance of the existing methods is still desired to improve. Also, despite the effectiveness of recent Transformer-based methods in medical image segmentation, the adaptability of Transformers is rarely investigated. In this paper, we present a novel UDA framework using a Transformer for building a cross-modality segmentation method with the advantages of learning long-range dependencies and transferring attentive information. To fully utilize the attention learned by the Transformer in UDA, we propose Meta Attention (MA) and use it to perform a fully attention-based alignment scheme, which can learn the hierarchical consistencies of attention and transfer more discriminative information between two modalities. We have conducted extensive experiments on cross-modality segmentation using three datasets, including a whole heart segmentation dataset (MMWHS), an abdominal organ segmentation dataset, and a brain tumor segmentation dataset. The promising results show that our method can significantly improve performance compared with the state-of-the-art UDA methods.

Learn2Reg: Comprehensive Multi-Task Medical Image Registration Challenge, Dataset and Evaluation in the Era of Deep Learning.

Image registration is a fundamental medical image analysis task, and a wide variety of approaches have been proposed. However, only a few studies have comprehensively compared medical image registration approaches on a wide range of clinically relevant tasks. This limits the development of registration methods, the adoption of research advances into practice, and a fair benchmark across competing approaches. The Learn2Reg challenge addresses these limitations by providing a multi-task medical image registration data set for comprehensive characterisation of deformable registration algorithms. A continuous evaluation will be possible at https://learn2reg.grand-challenge.org. Learn2Reg covers a wide range of anatomies (brain, abdomen, and thorax), modalities (ultrasound, CT, MR), availability of annotations, as well as intra- and inter-patient registration evaluation. We established an easily accessible framework for training and validation of 3D registration methods, which enabled the compilation of results of over 65 individual method submissions from more than 20 unique teams. We used a complementary set of metrics, including robustness, accuracy, plausibility, and runtime, enabling unique insight into the current state-of-the-art of medical image registration. This paper describes datasets, tasks, evaluation methods and results of the challenge, as well as results of further analysis of transferability to new datasets, the importance of label supervision, and resulting bias. While no single approach worked best across all tasks, many methodological aspects could be identified that push the performance of medical image registration to new state-of-the-art performance. Furthermore, we demystified the common belief that conventional registration methods have to be much slower than deep-learning-based methods.

Disease prediction with edge-variational graph convolutional networks.

The need for computational models that can incorporate imaging data with non-imaging data while investigating inter-subject associations arises in the task of population-based disease analysis. Although off-the-shelf deep convolutional neural networks have empowered representation learning from imaging data, incorporating data of different modalities complementarily in a unified model to improve the disease diagnostic quality is still challenging. In this work, we propose a generalizable graph-convolutional framework for population-based disease prediction on multi-modal medical data. Unlike previous methods constructing a static affinity population graph in a hand-crafting manner, the proposed framework can automatically learn to build a population graph with variational edges, which we show can be optimized jointly with spectral graph convolutional networks. In addition, to estimate the predictive uncertainty related to the constructed graph, we propose Monte-Carlo edge dropout uncertainty estimation. Experimental results on four multi-modal datasets demonstrate that the proposed method can substantially improve the predictive accuracy for Autism Spectrum Disorder, Alzheimer's disease, and ocular diseases. A sufficient ablation study with in-depth discussion is conducted to evaluate the effectiveness of each component and the choice of algorithmic details of the proposed method. The results indicate the potential and extendability of the proposed framework in leveraging multi-modal data for population-based disease prediction.

Relax and focus on brain tumor segmentation.

In this paper, we present a Deep Convolutional Neural Networks (CNNs) for fully automatic brain tumor segmentation for both high- and low-grade gliomas in MRI images. Unlike normal tissues or organs that usually have a fixed location or shape, brain tumors with different grades have shown great variation in terms of the location, size, structure, and morphological appearance. Moreover, the severe data imbalance exists not only between the brain tumor and non-tumor tissues, but also among the different sub-regions inside brain tumor (e.g., enhancing tumor, necrotic, edema, and non-enhancing tumor). Therefore, we introduce a hybrid model to address the challenges in the multi-modality multi-class brain tumor segmentation task. First, we propose the dynamic focal Dice loss function that is able to focus more on the smaller tumor sub-regions with more complex structures during training, and the learning capacity of the model is dynamically distributed to each class independently based on its training performance in different training stages. Besides, to better recognize the overall structure of the brain tumor and the morphological relationship among different tumor sub-regions, we relax the boundary constraints for the inner tumor regions in coarse-to-fine fashion. Additionally, a symmetric attention branch is proposed to highlight the possible location of the brain tumor from the asymmetric features caused by growth and expansion of the abnormal tissues in the brain. Generally, to balance the learning capacity of the model between spatial details and high-level morphological features, the proposed model relaxes the constraints of the inner boundary and complex details and enforces more attention on the tumor shape, location, and the harder classes of the tumor sub-regions. The proposed model is validated on the publicly available brain tumor dataset from real patients, BRATS 2019. The experimental results reveal that our model improves the overall segmentation performance in comparison with the state-of-the-art methods, with major progress on the recognition of the tumor shape, the structural relationship of tumor sub-regions, and the segmentation of more challenging tumor sub-regions, e.g., the tumor core, and enhancing tumor.

MedQ: Lossless ultra-low-bit neural network quantization for medical image segmentation.

Implementing deep convolutional neural networks (CNNs) with boolean arithmetic is ideal for eliminating the notoriously high computational expense of deep learning models. However, although lossless model compression via weight-only quantization has been achieved in previous works, it is still an open problem about how to reduce the computation precision of CNNs without losing performance, especially for medical image segmentation tasks where data dimension is high and annotation is scarce. This paper presents a novel CNN quantization framework that can squeeze a deep model (both parameters and activation) to extremely low bitwidth, e.g., 1∼2 bits, while maintaining its high performance. In the new method, we first design a strong baseline quantizer with an optimizable quantization range. Then, to relieve the back-propagation difficulty caused by the discontinuous quantization function, we design a radical residual connection scheme that allows gradients to flow through every quantized layer freely. Moreover, a tanh-based derivative function is used to further boost gradient flow and a distributional loss is employed to regularize the model output. Extensive experiments and ablation studies are conducted on two well-established public 3D segmentation datasets, i.e., BRATS2020 and LiTS. Experimental results evidence that our framework not only outperforms state-of-the-art quantization approaches significantly, but also achieves lossless performance on both datasets with ternary (2-bit) quantization.

3D Randomized Connection Network with Graph-based Label Inference.

In this paper, a novel 3D deep learning network is proposed for brain MR image segmentation with randomized connection, which can decrease the dependency between layers and increase the network capacity. The convolutional LSTM and 3D convolution are employed as network units to capture the long-term and short-term 3D properties respectively. To assemble these two kinds of spatial-temporal information and refine the deep learning outcomes, we further introduce an efficient graph-based node selection and label inference method. Experiments have been carried out on two publicly available databases and results demonstrate that the proposed method can obtain competitive performances as compared with other state-of-the-art methods.

Feature Sensitive Label Fusion With Random Walker for Atlas-Based Image Segmentation.

In this paper, a novel label fusion method is proposed for brain magnetic resonance image segmentation. This label fusion method is formulated on a graph, which embraces both label priors from atlases and anatomical priors from target image. To represent a pixel in a comprehensive way, three kinds of feature vectors are generated, including intensity, gradient, and structural signature. To select candidate atlas nodes for fusion, rather than exact searching, randomized k-d tree with spatial constraint is introduced as an efficient approximation for high-dimensional feature matching. Feature sensitive label prior (FSLP), which takes both the consistency and variety of different features into consideration, is proposed to gather atlas priors. As FSLP is a non-convex problem, one heuristic approach is further designed to solve it efficiently. Moreover, based on the anatomical knowledge, parts of the target pixels are also employed as the graph seeds to assist the label fusion process, and an iterative strategy is utilized to gradually update the label map. The comprehensive experiments carried out on two publicly available databases give results to demonstrate that the proposed method can obtain better segmentation quality.

A Markov Random Field Groupwise Registration Framework for Face Recognition.

In this paper, we propose a new framework for tackling face recognition problem. The face recognition problem is formulated as groupwise deformable image registration and feature matching problem. The main contributions of the proposed method lie in the following aspects: (1) Each pixel in a facial image is represented by an anatomical signature obtained from its corresponding most salient scale local region determined by the survival exponential entropy (SEE) information theoretic measure. (2) Based on the anatomical signature calculated from each pixel, a novel Markov random field based groupwise registration framework is proposed to formulate the face recognition problem as a feature guided deformable image registration problem. The similarity between different facial images are measured on the nonlinear Riemannian manifold based on the deformable transformations. (3) The proposed method does not suffer from the generalizability problem which exists commonly in learning based algorithms. The proposed method has been extensively evaluated on four publicly available databases: FERET, CAS-PEAL-R1, FRGC ver 2.0, and the LFW. It is also compared with several state-of-the-art face recognition approaches, and experimental results demonstrate that the proposed method consistently achieves the highest recognition rates among all the methods under comparison.

Label inference with registration and patch priors.

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

Nonrigid image registration with crystal dislocation energy.

The goal of nonrigid image registration is to find a suitable transformation such that the transformed moving image becomes similar to the reference image. The image registration problem can also be treated as an optimization problem, which tries to minimize an objective energy function that measures the differences between two involved images. In this paper, we consider image matching as the process of aligning object boundaries in two different images. The registration energy function can be defined based on the total energy associated with the object boundaries. The optimal transformation is obtained by finding the equilibrium state when the total energy is minimized, which indicates the object boundaries find their correspondences and stop deforming. We make an analogy between the above processes with the dislocation system in physics. The object boundaries are viewed as dislocations (line defects) in crystal. Then the well-developed dislocation energy is used to derive the energy assigned to object boundaries in images. The newly derived registration energy function takes the global gradient information of the entire image into consideration, and produces an orientation-dependent and long-range interaction between two images to drive the registration process. This property of interaction endows the new registration framework with both fast convergence rate and high registration accuracy. Moreover, the new energy function can be adapted to realize symmetric diffeomorphic transformation so as to ensure one-to-one matching between subjects. In this paper, the superiority of the new method is theoretically proven, experimentally tested and compared with the state-of-the-art SyN method. Experimental results with 3-D magnetic resonance brain images demonstrate that the proposed method outperforms the compared methods in terms of both registration accuracy and computation time.

Segmentation of intracranial vessels and aneurysms in phase contrast magnetic resonance angiography using multirange filters and local variances.

Segmentation of intensity varying and low-contrast structures is an extremely challenging and rewarding task. In computer-aided diagnosis of intracranial aneurysms, segmenting the high-intensity major vessels along with the attached low-contrast aneurysms is essential to the recognition of this lethal vascular disease. It is particularly helpful in performing early and noninvasive diagnosis of intracranial aneurysms using phase contrast magnetic resonance angiographic (PC-MRA) images. The major challenges of developing a PC-MRA-based segmentation method are the significantly varying voxel intensity inside vessels with different flow velocities and the signal loss in the aneurysmal regions where turbulent flows occur. This paper proposes a novel intensity-based algorithm to segment intracranial vessels and the attached aneurysms. The proposed method can handle intensity varying vasculatures and also the low-contrast aneurysmal regions affected by turbulent flows. It is grounded on the use of multirange filters and local variances to extract intensity-based image features for identifying contrast varying vasculatures. The extremely low-intensity region affected by turbulent flows is detected according to the topology of the structure detected by multirange filters and local variances. The proposed method is evaluated using a phantom image volume with an aneurysm and four clinical cases. It achieves 0.80 dice score in the phantom case. In addition, different components of the proposed method-the multirange filters, local variances, and topology-based detection-are evaluated in the comparison between the proposed method and its lower complexity variants. Owing to the analogy between these variants and existing vascular segmentation methods, this comparison also exemplifies the advantage of the proposed method over the existing approaches. It analyzes the weaknesses of these existing approaches and justifies the use of every component involved in the proposed method. It is shown that the proposed method is capable of segmenting blood vessels and the attached aneurysms on PC-MRA images.

Random walks with adaptive cylinder flux based connectivity for vessel segmentation.

In this paper, we present a novel graph-based method for segmenting the whole 3D vessel tree structures. Our method exploits a new adaptive cylinder flux (ACF) based connectivity framework, which is formulated based on random walks. To avoid the shrinking problem of elongated structure, all existing graph-based energy optimization methods for vessel segmentation rely on skeleton or ROI extraction. As a result, the performance of these vessel segmentation methods then depends heavily on the skeleton extraction results. In this paper, with the help of ACF based connectivity framework, a global optimal segmentation result can be obtained without extracting skeleton or ROI. The classical issues of the graph-based methods, such as shrinking bias and sensitivity to seed point location, can be solved effectively with the proposed method thanks to the connectivity framework.

Principal curves for lumen center extraction and flow channel width estimation in 3-D arterial networks: theory, algorithm, and validation.

We present an energy-minimization-based framework for locating the centerline and estimating the width of tubelike objects from their structural network with a nonparametric model. The nonparametric representation promotes simple modeling of nested branches and n -way furcations, i.e., structures that abound in an arterial network, e.g., a cerebrovascular circulation. Our method is capable of extracting the entire vascular tree from an angiogram in a single execution with a proper initialization. A succinct initial model from the user with arterial network inlets, outlets, and branching points is sufficient for complex vasculature. The novel method is based upon the theory of principal curves. In this paper, theoretical extension to grayscale angiography is discussed, and an algorithm to find an arterial network as principal curves is also described. Quantitative validation on a number of simulated data sets, synthetic volumes of 19 BrainWeb vascular models, and 32 Rotterdam Coronary Artery volumes was conducted. We compared the algorithm to a state-of-the-art method and further tested it on two clinical data sets. Our algorithmic outputs-lumen centers and flow channel widths-are important to various medical and clinical applications, e.g., vasculature segmentation, registration and visualization, virtual angioscopy, and vascular atlas formation and population study.

Nonrigid brain MR image registration using uniform spherical region descriptor.

There are two main issues that make nonrigid image registration a challenging task. First, voxel intensity similarity may not be necessarily equivalent to anatomical similarity in the image correspondence searching process. Second, during the imaging process, some interferences such as unexpected rotations of input volumes and monotonic gray-level bias fields can adversely affect the registration quality. In this paper, a new feature-based nonrigid image registration method is proposed. The proposed method is based on a new type of image feature, namely, uniform spherical region descriptor (USRD), as signatures for each voxel. The USRD is rotation and monotonic gray-level transformation invariant and can be efficiently calculated. The registration process is therefore formulated as a feature matching problem. The USRD feature is integrated with the Markov random field labeling framework in which energy function is defined for registration. The energy function is then optimized by the α-expansion algorithm. The proposed method has been compared with five state-of-the-art registration approaches on both the simulated and real 3-D databases obtained from the BrainWeb and Internet Brain Segmentation Repository, respectively. Experimental results demonstrate that the proposed method can achieve high registration accuracy and reliable robustness behavior.

Minimum average-cost path for real time 3D coronary artery segmentation of CT images.

In this paper, we propose a Minimum Average-cost Path (MACP) model for segmenting 3D coronary arteries by minimizing the average edge cost along path in discrete 4D graph constructed by image voxels and associated radii. Prim's Minimum Spanning Tree method is used for efficient optimization of the MACP model. The centerline and the radii of the cross sections of the coronary artery are extracted simultaneously during the optimization. The method does not need any image preprocessing steps and has been intensively validated as an effective approach with the Rotterdam Coronary Artery Algorithm Evaluation Framework. The computational cost of the proposed method is particularly low (7.467 seconds per segment, 18.5mm/s on average), which makes real time segmentation of coronary artery possible. Shortcut problem, which is a classic issue of the minimal path techniques, can also be overcome by the proposed method.

VE-LLI-VO: vessel enhancement using local line integrals and variational optimization.

Vessel enhancement is a primary preprocessing step for vessel segmentation and visualization of vasculatures. In this paper, a new vessel enhancement technique is proposed in order to produce accurate vesselness measures and vessel direction estimations that are less subject to local intensity abnormalities. The proposed method is called vessel enhancement using local line integrals and variational optimization (VE-LLI-VO). First, vessel enhancement using local line integrals (VE-LLI) is introduced in which a vessel model is embedded by regarding a vessel segment as a straight line based upon the second order information of the local line integrals. Useful quantities similar to the eigenvalues and eigenvectors of the Hessian matrix are produced. Moreover, based upon the local line integrals, junctions can be detected and handled effectively. This can help deal with the bifurcation suppression problem which exists in the Hessian-based enhancement methods. Then a more generic curve model is embedded to model vessels and a variational optimization (VO) framework is introduced to generate optimized vesselness measures. Experiments have been conducted on both synthetic images and retinal images. It is experimentally demonstrated that VE-LLI-VO produces improved performance as compared with the widely used techniques in terms of both vesselness measurement and vessel direction estimation.

Learning-based multi-modal rigid image registration by using Bhattacharyya distances.

Multi-modal image registration is a momentous technology in medical image processing and analysis. In order to improve the robustness and accuracy of multi-modal rigid image registration, a novel learning-based dissimilarity function is proposed in this paper. This novel dissimilarity function is based on measuring the dissimilarity between the joint intensity distribution of the testing image pair and the expected intensity distributions, which is learned from a registered image pair, with Bhattacharyya distances. Then, the aim of the registration process is to minimize the dissimilarity function. Eight hundred randomized CT - T1 registrations were performed and evaluated by the Retrospective Image Registration Evaluation (RIRE) project. The experimental results demonstrate that the proposed method can achieve higher robustness and accuracy, as compared with a closely related approach and a state-of-the-art method.

Feature based nonrigid brain MR image registration with symmetric alpha stable filters.

A new feature based nonrigid image registration method for magnetic resonance (MR) brain images is presented in this paper. Each image voxel is represented by a rotation invariant feature vector, which is computed by passing the input image volumes through a new bank of symmetric alpha stable (SalphaS) filters. There are three main contributions presented in this paper. First, this work is motivated by the fact that the frequency spectrums of the brain MR images often exhibit non-Gaussian heavy-tail behavior which cannot be satisfactorily modeled by the conventional Gabor filters. To this end, we propose the use of SalphaS filters to model such behavior and show that the Gabor filter is a special case of the SalphaS filter. Second, the maximum response orientation (MRO) selection criterion is designed to extract rotation invariant features for registration tasks. The MRO selection criterion also significantly reduces the number of dimensions of feature vectors and therefore lowers the computation time. Third, in case the segmentations of the input image volumes are available, the Fisher's separation criterion (FSC) is introduced such that the discriminating power of different feature types can be directly compared with each other before performing the registration process. Using FSC, weights can also be assigned automatically to different voxels in the brain MR images. The weight of each voxel determined by FSC reflects how distinctive and salient the voxel is. Using the most distinctive and salient voxels at the initial stage to drive the registration can reduce the risk of being trapped in the local optimum during image registration process. The larger the weight, the more important the voxel. With the extracted feature vectors and the associated weights, the proposed method registers the source and the target images in a hierarchical multiresolution manner. The proposed method has been intensively evaluated on both simulated and real 3-D datasets obtained from BrainWeb and Internet Brain Segmentation Repository (IBSR), respectively, and compared with HAMMER, an extended version of HAMMER based on local histograms (LHF), FFD, Demons, and the Gabor filter based registration method. It is shown that the proposed method achieves the highest registration accuracy among the five widely used image registration methods.