MMUDA: A Mutual Multi-Task Learning Network with Unsupervised Domain Adaptation for Glaucoma Screening

Glaucoma is a leading cause of irreversible blindness worldwide, requiring early screening for timely treatment to maintain vision and quality of life. The automated multi-task screening approaches conduct glaucoma classification, as well as optic cup and optic disc segmentation simultaneously, enhancing interpretability by providing probabilities and essential morphological biomarkers. However, existing methods struggle with generalization, as they fail to adapt to the domain shift encountered across different datasets on both tasks in parallel. To address this issue, we propose a novel mutual multi-task network with unsupervised domain adaptation (MMUDA) to realize cross-domain accurate glaucoma screening. We devise a multi-scale classification branch for glaucoma and an adversarial segmentation branch for optic cup and optic disk, ensuring global domain alignment. Additionally, we design a multi-task mutual pseudo supervision (MMPS) training strategy for aligning category-wise domain representation in both tasks. Furthermore, our network presents the dual dimensional self-attention (DDSA) mechanism to enhance cross-domain relationship understanding within mixed-domain mini-batches. Extensive experiments on three public datasets prove that MMUDA outperforms the unsupervised domain adaptation methods in glaucoma screening. Our source code is available at https://github.com/AllenYT/MMUDA.