Convolutional neural networks have been widely applied to medical image segmentation and have achieved considerable performance. However, the performance may be significantly affected by the domain gap between training data (source domain) and testing data (target domain). To address this issue, we propose a data manipulation based domain generalization method, called Automated Augmentation for Domain Generalization (AADG). Our AADG framework can effectively sample data augmentation policies that generate novel domains and diversify the training set from an appropriate search space. Specifically, we introduce a novel proxy task maximizing the diversity among multiple augmented novel domains as measured by the Sinkhorn distance in a unit sphere space, making automated augmentation tractable. Adversarial training and deep reinforcement learning are employed to efficiently search the objectives. Quantitative and qualitative experiments on 11 publicly-accessible fundus image datasets (four for retinal vessel segmentation, four for optic disc and cup (OD/OC) segmentation and three for retinal lesion segmentation) are comprehensively performed. Two OCTA datasets for retinal vasculature segmentation are further involved to validate cross-modality generalization. Our proposed AADG exhibits state-of-the-art generalization performance and outperforms existing approaches by considerable margins on retinal vessel, OD/OC and lesion segmentation tasks. The learned policies are empirically validated to be model-agnostic and can transfer well to other models. The source code is available at https://github.com/CRazorback/AADG.

Stroke is a serious manifestation of various cerebrovascular diseases and one of the most dangerous diseases in the world today. Volume quantification and location detection of chronic stroke lesions provide vital biomarkers for stroke rehabilitation. Recently, deep learning has seen a rapid growth, with a great potential in segmenting medical images. In this work, unlike most deep learning-based segmentation methods utilizing only magnetic resonance (MR) images as the input, we propose and validate a novel stroke lesion segmentation approach named multi-inputs UNet (MI-UNet) that incorporates brain parcellation information, including gray matter (GM), white matter (WM) and lateral ventricle (LV). The brain parcellation is obtained from 3D diffeomorphic registration and is concatenated with the original MR image to form two-channel inputs to the subsequent MI-UNet. Effectiveness of the proposed pipeline is validated using a dataset consisting of 229 T1-weighted MR images. Experiments are conducted via a five-fold cross-validation. The proposed MI-UNet performed significantly better than UNet in both 2D and 3D settings. Our best results obtained by 3D MI-UNet has superior segmentation performance, as measured by the Dice score, Hausdorff distance, average symmetric surface distance, as well as precision, over other state-of-the-art methods.

Automated detection of exudates from fundus images plays an important role in diabetic retinopathy (DR) screening and evaluation, for which supervised or semi-supervised learning methods are typically preferred. However, a potential limitation of supervised and semi-supervised learning based detection algorithms is that they depend substantially on the sample size of training data and the quality of annotations, which is the fundamental motivation of this work. In this study, we construct a dataset containing 1219 fundus images (from DR patients and healthy controls) with annotations of exudate lesions. In addition to exudate annotations, we also provide four additional labels for each image: left-versus-right eye label, DR grade (severity scale) from three different grading protocols, the bounding box of the optic disc (OD), and fovea location. This dataset provides a great opportunity to analyze the accuracy and reliability of different exudate detection, OD detection, fovea localization, and DR classification algorithms. Moreover, it will facilitate the development of such algorithms in the realm of supervised and semi-supervised learning.

Although deep learning-based diabetic retinopathy (DR) classification methods typically benefit from well-designed architectures of convolutional neural networks, the training setting also has a non-negligible impact on prediction performance. The training setting includes various interdependent components, such as an objective function, a data sampling strategy, and a data augmentation approach. To identify the key components in a standard deep learning framework (ResNet-50) for DR grading, we systematically analyze the impact of several major components. Extensive experiments are conducted on a publicly available dataset EyePACS. We demonstrate that (1) the DR grading framework is sensitive to input resolution, objective function, and composition of data augmentation; (2) using mean square error as the loss function can effectively improve the performance with respect to a task-specific evaluation metric, namely the quadratically weighted Kappa; (3) utilizing eye pairs boosts the performance of DR grading and; (4) using data resampling to address the problem of imbalanced data distribution in EyePACS hurts the performance. Based on these observations and an optimal combination of the investigated components, our framework, without any specialized network design, achieves a state-of-the-art result (0.8631 for Kappa) on the EyePACS test set (a total of 42,670 fundus images) with only image-level labels. We also examine the proposed training practices on other fundus datasets and other network architectures to evaluate their generalizability. Our codes and pre-trained model are available online.

Self-supervised Learning (SSL) has been widely applied to learn image representations through exploiting unlabeled images. However, it has not been fully explored in the medical image analysis field. In this work, Saliency-guided Self-Supervised image Transformer (SSiT) is proposed for Diabetic Retinopathy (DR) grading from fundus images. We novelly introduce saliency maps into SSL, with a goal of guiding self-supervised pre-training with domain-specific prior knowledge. Specifically, two saliency-guided learning tasks are employed in SSiT: (1) Saliency-guided contrastive learning is conducted based on the momentum contrast, wherein fundus images' saliency maps are utilized to remove trivial patches from the input sequences of the momentum-updated key encoder. Thus, the key encoder is constrained to provide target representations focusing on salient regions, guiding the query encoder to capture salient features. (2) The query encoder is trained to predict the saliency segmentation, encouraging the preservation of fine-grained information in the learned representations. To assess our proposed method, four publicly-accessible fundus image datasets are adopted. One dataset is employed for pre-training, while the three others are used to evaluate the pre-trained models' performance on downstream DR grading. The proposed SSiT significantly outperforms other representative state-of-the-art SSL methods on all downstream datasets and under various evaluation settings. For example, SSiT achieves a Kappa score of 81.88% on the DDR dataset under fine-tuning evaluation, outperforming all other ViT-based SSL methods by at least 9.48%.