UAU-Net: United Attention U-Shaped Network for the Segmentation of Pigment Deposits in Fundus Images of Retinitis Pigmentosa

Retinitis Pigmentosa (RP) is a retinal disease with high rate of blindness. Retinal pigment deposits are a typical symptom of RP, whose automatic segmentation is crucial to the early diagnosis of RP. In fundus images, pigment deposits have various shapes and sizes and are scattered randomly, which makes the automatic segmentation very challenging. In this paper, we propose a United Attention U-shaped Network (UAU-Net) for segmentation of pigment deposits in fundus images, comprising four parts: encoder, decoder, Multi-scale Global Attention Module (MsGAM), and Spatial-enhanced Attention Module (SEAM). The MsGAM is proposed to extract multi-scale spatial and channel information in constructing associations between different locations of pigment deposits, and the SEAM is proposed to preserve detailed features and enhance the model's ability to segment small targets. Comprehensive experiments on 215 fundus images show that UAU-Net outperforms other state-of-the-art methods with Dice and Intersection-over-Union of 60.25% and 44.91%, respectively.