Automated lesion segmentation in fundus images with many-to-many reassembly of features

Existing CNN-based segmentation approaches have achieved remarkable progresses on segmenting objects in regular sizes. However, when migrating them to segment tiny retinal lesions, they encounter challenges. The feature reassembly operators that they adopt are prone to discard the subtle activations about tiny lesions and fail to capture long-term dependencies. This paper aims to solve these issues and proposes a novel Many-to-Many Reassembly of Features (M2MRF) for tiny lesion segmentation. Our proposed M2MRF reassembles features in a dimension-reduced feature space and simultaneously aggregates multiple features inside a large predefined region into multiple output features. In this way, subtle activations about small lesions can be maintained as much as possible and long-term spatial dependencies can be captured to further enhance the lesion features. Experimental results on two lesion segmentation benchmarks, i.e., DDR and IDRiD, show that 1) our M2MRF outperforms existing feature reassembly operators, and 2) equipped with our M2MRF, the HRNetV2 is able to achieve substantially better performances and generalisation ability than existing methods. Our code is made publicly available at https://github.com/CVIU-CSU/M2MRF-Lesion-Segmentation.