PSScreen V2: Partially Supervised Multiple Retinal Disease Screening

TL;DR

PSScreen V2 tackles multi-disease screening under partial labeling and domain shift by integrating a teacher–student three-branch framework with two novel frequency-domain augmentations, LF-Dropout and LF-Uncert. A text-guided semantic decoupling module helps focus on disease-relevant regions, while pseudo labeling enables learning from unknown classes. Extensive retinal and chest X-ray experiments show state-of-the-art in-domain and out-of-domain performance, strong backbone compatibility (including DINOv2), and robust ablation results validating each component. The approach offers a scalable, data-efficient path for generalizable medical image screening across diverse datasets and modalities.

Abstract

In this work, we propose PSScreen V2, a partially supervised self-training framework for multiple retinal disease screening. Unlike previous methods that rely on fully labelled or single-domain datasets, PSScreen V2 is designed to learn from multiple partially labelled datasets with different distributions, addressing both label absence and domain shift challenges. To this end, PSScreen V2 adopts a three-branch architecture with one teacher and two student networks. The teacher branch generates pseudo labels from weakly augmented images to address missing labels, while the two student branches introduce novel feature augmentation strategies: Low-Frequency Dropout (LF-Dropout), which enhances domain robustness by randomly discarding domain-related low-frequency components, and Low-Frequency Uncertainty (LF-Uncert), which estimates uncertain domain variability via adversarially learned Gaussian perturbations of low-frequency statistics. Extensive experiments on multiple in-domain and out-of-domain fundus datasets demonstrate that PSScreen V2 achieves state-of-the-art performance and superior domain generalization ability. Furthermore, compatibility tests with diverse backbones, including the vision foundation model DINOv2, as well as evaluations on chest X-ray datasets, highlight the universality and adaptability of the proposed framework. The codes are available at https://github.com/boyiZheng99/PSScreen_V2.

Figures (7)

• Figure 1: Examples of open-access datasets for retinal disease screening and model comparisons under the three learning paradigms. (a) lists open-access datasets where "✓" indicates labels for diseases are available while "?" denotes labels are not available. From (b) to (d), we illustrate the pipelines and characteristics of the fully supervised screening model usually trained with a fully labelled dataset, the self-supervised screening model trained with large-scale image-text pairs, and the partially supervised screening model trained with multiple partially labelled datasets.
• Figure 2: (a) Illustration of the PSScreen V2. PSScreen V2 adopts a three-branch self-training framework, where two student branches are supervised by pseudo labels from a common $\mathsf{Teacher}$$\mathcal{T}$. $\mathsf{Student1}$$\mathcal{S}_{1}$ augments features via low-frequency dropout (LF-Dropout), while $\mathsf{Student2}$$\mathcal{S}_{2}$ augments features via low-frequency uncertainty (LF-Uncert). All three branches share a backbone, a text-guided semantic decoupling module and a multi-label classifier. (b) An illustration of text-guided semantic decoupling module. (c) An illustration of LF-Uncert.
• Figure 3: Performance comparison of zero-shot inference with foundation models on the ODIR200$\times$3 dataset.
• Figure 4: $mQWK$ under varying augmentation range $r$.
• Figure 5: $mQWK$ under varying dropout probability $p$.