Self-training adversarial learning for cross-domain retinal OCT fluid segmentation

Contrastive Learning vs. Self-Learning vs. Deformable Data Augmentation in Semantic Segmentation of Medical Images

To develop a robust segmentation model, encoding the underlying features/structures of the input data is essential to discriminate the target structure from the background. To enrich the extracted feature maps, contrastive learning and self-learning techniques are employed, particularly when the size of the training dataset is limited. In this work, we set out to investigate the impact of contrastive learning and self-learning on the performance of the deep learning-based semantic segmentation. To this end, three different datasets were employed used for brain tumor and hippocampus delineation from MR images (BraTS and Decathlon datasets, respectively) and kidney segmentation from CT images (Decathlon dataset). Since data augmentation techniques are also aimed at enhancing the performance of deep learning methods, a deformable data augmentation technique was proposed and compared with contrastive learning and self-learning frameworks. The segmentation accuracy for the three datasets was assessed with and without applying data augmentation, contrastive learning, and self-learning to individually investigate the impact of these techniques. The self-learning and deformable data augmentation techniques exhibited comparable performance with Dice indices of 0.913 ± 0.030 and 0.920 ± 0.022 for kidney segmentation, 0.890 ± 0.035 and 0.898 ± 0.027 for hippocampus segmentation, and 0.891 ± 0.045 and 0.897 ± 0.040 for lesion segmentation, respectively. These two approaches significantly outperformed the contrastive learning and the original model with Dice indices of 0.871 ± 0.039 and 0.868 ± 0.042 for kidney segmentation, 0.872 ± 0.045 and 0.865 ± 0.048 for hippocampus segmentation, and 0.870 ± 0.049 and 0.860 ± 0.058 for lesion segmentation, respectively. The combination of self-learning with deformable data augmentation led to a robust segmentation model with no outliers in the outcomes. This work demonstrated the beneficial impact of self-learning and deformable data augmentation on organ and lesion segmentation, where no additional training datasets are needed.

Semi-supervised 3D retinal fluid segmentation via correlation mutual learning with global reasoning attention

Accurate 3D segmentation of fluid lesions in optical coherence tomography (OCT) is crucial for the early diagnosis of diabetic macular edema (DME). However, higher-dimensional spatial complexity and limited annotated data present significant challenges for effective 3D lesion segmentation. To address these issues, we propose a novel semi-supervised strategy using a correlation mutual learning framework for segmenting 3D DME lesions from 3D OCT images. Our method integrates three key innovations: (1) a shared encoder with three parallel, slightly different decoders, exhibiting cognitive biases and calculating statistical discrepancies among the decoders to represent uncertainty in unlabeled challenging regions. (2) a global reasoning attention module integrated into the encoder's output to transfer label prior knowledge to unlabeled data; and (3) a correlation mutual learning scheme, enforcing mutual consistency between one decoder's probability map and the soft pseudo labels generated by the other decoders. Extensive experiments demonstrate that our approach outperforms state-of-the-art (SOTA) methods, highlighting the potential of our framework for tackling the complex task of 3D retinal lesion segmentation.

Automatic height measurement of central serous chorioretinopathy lesion using a deep learning and adaptive gradient threshold based cascading strategy

Accurately quantifying the height of central serous chorioretinopathy (CSCR) lesion is of great significance for assisting ophthalmologists in diagnosing CSCR and evaluating treatment efficacy. The manual measurement results dominated by single optical coherence tomography (OCT) B-scan image in clinical practice face the dilemma of weak reference, poor reproducibility, and experience dependence. In this context, this paper constructs two schemes: Scheme Ⅰ draws on the idea of ensemble learning, namely, integrating multiple models for locating starting key point in the height direction of lesion in the inference stage, which appropriately improves the performance of a single model. Scheme Ⅱ designs an adaptive gradient threshold (AGT) technique, followed by the construction of cascading strategy, which involves preliminary location of starting key point through deep learning, and then employs AGT for precise adjustment. This strategy not only achieves effective location for starting key point, but also significantly reduces the large appetite of deep learning model for training samples. Subsequently, AGT continues to play a crucial role in locating the terminal key point in the height direction of lesion, further demonstrating its feasibility and effectiveness. Quantitative and qualitative key point location experiments in the height direction of lesion on 1152 samples, as well as the final height measurement display, consistently conveys the superiority of the constructed schemes, especially the cascading strategy, expanding another potential tool for the comprehensive analysis of CSCR.

Deep learning for unsupervised domain adaptation in medical imaging: Recent advancements and future perspectives

Deep learning has demonstrated remarkable performance across various tasks in medical imaging. However, these approaches primarily focus on supervised learning, assuming that the training and testing data are drawn from the same distribution. Unfortunately, this assumption may not always hold true in practice. To address these issues, unsupervised domain adaptation (UDA) techniques have been developed to transfer knowledge from a labeled domain to a related but unlabeled domain. In recent years, significant advancements have been made in UDA, resulting in a wide range of methodologies, including feature alignment, image translation, self-supervision, and disentangled representation methods, among others. In this paper, we provide a comprehensive literature review of recent deep UDA approaches in medical imaging from a technical perspective. Specifically, we categorize current UDA research in medical imaging into six groups and further divide them into finer subcategories based on the different tasks they perform. We also discuss the respective datasets used in the studies to assess the divergence between the different domains. Finally, we discuss emerging areas and provide insights and discussions on future research directions to conclude this survey.

Semi-supervised contrast learning-based segmentation of choroidal vessel in optical coherence tomography images

Objective.Choroidal vessels account for 85% of all blood vessels in the eye, and the accurate segmentation of choroidal vessels from optical coherence tomography (OCT) images provides important support for the quantitative analysis of choroid-related diseases and the development of treatment plans. Although deep learning-based methods have great potential for segmentation, these methods rely on large amounts of well-labeled data, and the data collection process is both time-consuming and laborious.Approach.In this paper, we propose a novel asymmetric semi-supervised segmentation framework called SSCR, based on a student-teacher model, to segment choroidal vessels in OCT images. The proposed framework enhances the segmentation results with uncertainty-aware self-integration and transformation consistency techniques. Meanwhile, we designed an asymmetric encoder-decoder network called Pyramid Pooling SegFormer (APP-SFR) for choroidal vascular segmentation. The network combines local attention and global attention information to improve the model's ability to learn complex vascular features. Additionally, we proposed a boundary repair module that enhances boundary confidence by utilizing a repair head to re-predict selected fuzzy points and further refines the segmentation boundary.Main results.We conducted extensive experiments on three different datasets: the ChorVessel dataset with 400 OCT images, the Meibomian Glands (MG) dataset with 400 images, and the U2OS Cell Nucleus Dataset with 200 images. The proposed method achieved an average Dice score of 74.23% on the ChorVessel dataset, which is 2.95% higher than the fully supervised network (U-Net) and outperformed other comparison methods. In both the MG dataset and the U2OS cell nucleus dataset, our proposed SSCR method achieved average Dice scores of 80.10% and 87.26%, respectively.Significance.The experimental results show that our proposed methods achieve better segmentation accuracy than other state-of-the-art methods. The method is designed to help clinicians make rapid diagnoses of ophthalmic diseases and has potential for clinical application.