Although Transformers have successfully transitioned from their language modelling origins to image-based applications, their quadratic computational complexity remains a challenge, particularly for dense prediction. In this paper we propose a content-based sparse attention method, as an alternative to dense self-attention, aiming to reduce the computation complexity while retaining the ability to model long-range dependencies. Specifically, we cluster and then aggregate key and value tokens, as a content-based method of reducing the total token count. The resulting clustered-token sequence retains the semantic diversity of the original signal, but can be processed at a lower computational cost. Besides, we further extend the clustering-guided attention from single-scale to multi-scale, which is conducive to dense prediction tasks. We label the proposed Transformer architecture ClusTR, and demonstrate that it achieves state-of-the-art performance on various vision tasks but at lower computational cost and with fewer parameters. For instance, our ClusTR small model with 22.7M parameters achieves 83.2\% Top-1 accuracy on ImageNet. Source code and ImageNet models will be made publicly available.
Multi-modal learning focuses on training models by equally combining multiple input data modalities during the prediction process. However, this equal combination can be detrimental to the prediction accuracy because different modalities are usually accompanied by varying levels of uncertainty. Using such uncertainty to combine modalities has been studied by a couple of approaches, but with limited success because these approaches are either designed to deal with specific classification or segmentation problems and cannot be easily translated into other tasks, or suffer from numerical instabilities. In this paper, we propose a new Uncertainty-aware Multi-modal Learner that estimates uncertainty by measuring feature density via Cross-modal Random Network Prediction (CRNP). CRNP is designed to require little adaptation to translate between different prediction tasks, while having a stable training process. From a technical point of view, CRNP is the first approach to explore random network prediction to estimate uncertainty and to combine multi-modal data. Experiments on two 3D multi-modal medical image segmentation tasks and three 2D multi-modal computer vision classification tasks show the effectiveness, adaptability and robustness of CRNP. Also, we provide an extensive discussion on different fusion functions and visualization to validate the proposed model.
Assessment of myocardial viability is essential in diagnosis and treatment management of patients suffering from myocardial infarction, and classification of pathology on myocardium is the key to this assessment. This work defines a new task of medical image analysis, i.e., to perform myocardial pathology segmentation (MyoPS) combining three-sequence cardiac magnetic resonance (CMR) images, which was first proposed in the MyoPS challenge, in conjunction with MICCAI 2020. The challenge provided 45 paired and pre-aligned CMR images, allowing algorithms to combine the complementary information from the three CMR sequences for pathology segmentation. In this article, we provide details of the challenge, survey the works from fifteen participants and interpret their methods according to five aspects, i.e., preprocessing, data augmentation, learning strategy, model architecture and post-processing. In addition, we analyze the results with respect to different factors, in order to examine the key obstacles and explore potential of solutions, as well as to provide a benchmark for future research. We conclude that while promising results have been reported, the research is still in the early stage, and more in-depth exploration is needed before a successful application to the clinics. Note that MyoPS data and evaluation tool continue to be publicly available upon registration via its homepage (www.sdspeople.fudan.edu.cn/zhuangxiahai/0/myops20/).
Self-supervised learning (SSL) opens up huge opportunities for better utilizing unlabeled data. It is essential for medical image analysis that is generally known for its lack of annotations. However, when we attempt to use as many as possible unlabeled medical images in SSL, breaking the dimension barrier (\ie, making it possible to jointly use both 2D and 3D images) becomes a must. In this paper, we propose a Universal Self-Supervised Transformer (USST) framework based on the student-teacher paradigm, aiming to leverage a huge of unlabeled medical data with multiple dimensions to learn rich representations. To achieve this, we design a Pyramid Transformer U-Net (PTU) as the backbone, which is composed of switchable patch embedding (SPE) layers and Transformer layers. The SPE layer switches to either 2D or 3D patch embedding depending on the input dimension. After that, the images are converted to a sequence regardless of their original dimensions. The Transformer layer then models the long-term dependencies in a sequence-to-sequence manner, thus enabling USST to learn representations from both 2D and 3D images. USST has two obvious merits compared to current dimension-specific SSL: (1) \textbf{more effective} - can learn representations from more and diverse data; and (2) \textbf{more versatile} - can be transferred to various downstream tasks. The results show that USST provides promising results on six 2D/3D medical image classification and segmentation tasks, outperforming the supervised ImageNet pre-training and advanced SSL counterparts substantially.
The domain gap caused mainly by variable medical image quality renders a major obstacle on the path between training a segmentation model in the lab and applying the trained model to unseen clinical data. To address this issue, domain generalization methods have been proposed, which however usually use static convolutions and are less flexible. In this paper, we propose a multi-source domain generalization model, namely domain and content adaptive convolution (DCAC), for medical image segmentation. Specifically, we design the domain adaptive convolution (DAC) module and content adaptive convolution (CAC) module and incorporate both into an encoder-decoder backbone. In the DAC module, a dynamic convolutional head is conditioned on the predicted domain code of the input to make our model adapt to the unseen target domain. In the CAC module, a dynamic convolutional head is conditioned on the global image features to make our model adapt to the test image. We evaluated the DCAC model against the baseline and four state-of-the-art domain generalization methods on the prostate segmentation, COVID-19 lesion segmentation, and optic cup/optic disc segmentation tasks. Our results indicate that the proposed DCAC model outperforms all competing methods on each segmentation task, and also demonstrate the effectiveness of the DAC and CAC modules.
Convolutional neural networks (CNNs) have been the de facto standard for nowadays 3D medical image segmentation. The convolutional operations used in these networks, however, inevitably have limitations in modeling the long-range dependency due to their inductive bias of locality and weight sharing. Although Transformer was born to address this issue, it suffers from extreme computational and spatial complexities in processing high-resolution 3D feature maps. In this paper, we propose a novel framework that efficiently bridges a {\bf Co}nvolutional neural network and a {\bf Tr}ansformer {\bf (CoTr)} for accurate 3D medical image segmentation. Under this framework, the CNN is constructed to extract feature representations and an efficient deformable Transformer (DeTrans) is built to model the long-range dependency on the extracted feature maps. Different from the vanilla Transformer which treats all image positions equally, our DeTrans pays attention only to a small set of key positions by introducing the deformable self-attention mechanism. Thus, the computational and spatial complexities of DeTrans have been greatly reduced, making it possible to process the multi-scale and high-resolution feature maps, which are usually of paramount importance for image segmentation. We conduct an extensive evaluation on the Multi-Atlas Labeling Beyond the Cranial Vault (BCV) dataset that covers 11 major human organs. The results indicate that our CoTr leads to a substantial performance improvement over other CNN-based, transformer-based, and hybrid methods on the 3D multi-organ segmentation task. Code is available at \def\UrlFont{\rm\small\ttfamily} \url{https://github.com/YtongXie/CoTr}
Automated and accurate 3D medical image segmentation plays an essential role in assisting medical professionals to evaluate disease progresses and make fast therapeutic schedules. Although deep convolutional neural networks (DCNNs) have widely applied to this task, the accuracy of these models still need to be further improved mainly due to their limited ability to 3D context perception. In this paper, we propose the 3D context residual network (ConResNet) for the accurate segmentation of 3D medical images. This model consists of an encoder, a segmentation decoder, and a context residual decoder. We design the context residual module and use it to bridge both decoders at each scale. Each context residual module contains both context residual mapping and context attention mapping, the formal aims to explicitly learn the inter-slice context information and the latter uses such context as a kind of attention to boost the segmentation accuracy. We evaluated this model on the MICCAI 2018 Brain Tumor Segmentation (BraTS) dataset and NIH Pancreas Segmentation (Pancreas-CT) dataset. Our results not only demonstrate the effectiveness of the proposed 3D context residual learning scheme but also indicate that the proposed ConResNet is more accurate than six top-ranking methods in brain tumor segmentation and seven top-ranking methods in pancreas segmentation. Code is available at https://git.io/ConResNet
It has been widely recognized that the success of deep learning in image segmentation relies overwhelmingly on a myriad amount of densely annotated training data, which, however, are difficult to obtain due to the tremendous labor and expertise required, particularly for annotating 3D medical images. Although self-supervised learning (SSL) has shown great potential to address this issue, most SSL approaches focus only on image-level global consistency, but ignore the local consistency which plays a pivotal role in capturing structural information for dense prediction tasks such as segmentation. In this paper, we propose a PriorGuided Local (PGL) self-supervised model that learns the region-wise local consistency in the latent feature space. Specifically, we use the spatial transformations, which produce different augmented views of the same image, as a prior to deduce the location relation between two views, which is then used to align the feature maps of the same local region but being extracted on two views. Next, we construct a local consistency loss to minimize the voxel-wise discrepancy between the aligned feature maps. Thus, our PGL model learns the distinctive representations of local regions, and hence is able to retain structural information. This ability is conducive to downstream segmentation tasks. We conducted an extensive evaluation on four public computerized tomography (CT) datasets that cover 11 kinds of major human organs and two tumors. The results indicate that using pre-trained PGL model to initialize a downstream network leads to a substantial performance improvement over both random initialization and the initialization with global consistency-based models. Code and pre-trained weights will be made available at: https://git.io/PGL.
Due to the intensive cost of labor and expertise in annotating 3D medical images at a voxel level, most benchmark datasets are equipped with the annotations of only one type of organs and/or tumors, resulting in the so-called partially labeling issue. To address this, we propose a dynamic on-demand network (DoDNet) that learns to segment multiple organs and tumors on partially labeled datasets. DoDNet consists of a shared encoder-decoder architecture, a task encoding module, a controller for generating dynamic convolution filters, and a single but dynamic segmentation head. The information of the current segmentation task is encoded as a task-aware prior to tell the model what the task is expected to solve. Different from existing approaches which fix kernels after training, the kernels in dynamic head are generated adaptively by the controller, conditioned on both input image and assigned task. Thus, DoDNet is able to segment multiple organs and tumors, as done by multiple networks or a multi-head network, in a much efficient and flexible manner. We have created a large-scale partially labeled dataset, termed MOTS, and demonstrated the superior performance of our DoDNet over other competitors on seven organ and tumor segmentation tasks. We also transferred the weights pre-trained on MOTS to a downstream multi-organ segmentation task and achieved state-of-the-art performance. This study provides a general 3D medical image segmentation model that has been pre-trained on a large-scale partially labelled dataset and can be extended (after fine-tuning) to downstream volumetric medical data segmentation tasks. The dataset and code areavailableat: https://git.io/DoDNet
Accurate and automated gland segmentation on histology tissue images is an essential but challenging task in the computer-aided diagnosis of adenocarcinoma. Despite their prevalence, deep learning models always require a myriad number of densely annotated training images, which are difficult to obtain due to extensive labor and associated expert costs related to histology image annotations. In this paper, we propose the pairwise relation-based semi-supervised (PRS^2) model for gland segmentation on histology images. This model consists of a segmentation network (S-Net) and a pairwise relation network (PR-Net). The S-Net is trained on labeled data for segmentation, and PR-Net is trained on both labeled and unlabeled data in an unsupervised way to enhance its image representation ability via exploiting the semantic consistency between each pair of images in the feature space. Since both networks share their encoders, the image representation ability learned by PR-Net can be transferred to S-Net to improve its segmentation performance. We also design the object-level Dice loss to address the issues caused by touching glands and combine it with other two loss functions for S-Net. We evaluated our model against five recent methods on the GlaS dataset and three recent methods on the CRAG dataset. Our results not only demonstrate the effectiveness of the proposed PR-Net and object-level Dice loss, but also indicate that our PRS^2 model achieves the state-of-the-art gland segmentation performance on both benchmarks.