RMCA U-net: Hard exudates segmentation for retinal fundus images

• Published in: Expert systems with applications 2023-12, Vol.234, p.120987, Article 120987
• Main Authors: Fu, Yinghua, Zhang, Ge, Lu, Xin, Wu, Honghan, Zhang, Dawei
• Format: Article
• Language: English
• Subjects: Channel attention; Hard exudate; Multi-scale feature fusion module; Residual module; RMCA U-net
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2023.120987

Hard exudate plays an important role in grading diabetic retinopathy (DR) as a critical indicator. Therefore, the accurate segmentation of hard exudates is of clinical importance. However, the percentage of hard exudates in the whole fundus image is relatively small, and their shapes are often irregular and the contrasts are usually not high enough. Hence, they are prone to misclassifications e.g., misclassified as part of the optic disc structure or cotton wool spots, which results in the low segmentation accuracy and efficiency. This paper proposes a novel neural network RMCA U-net to accurately segmentation hard exudate in fundus images. The network features a U-shape framework combined with a residual structure to obtain the subtle features of hard exudate. A multi-scale feature fusion (MSFF) module and an improved channel attention (CA) module are designed and involved to effectively segmentation sparse small lesions. The proposed method in this paper has been trained and evaluated on three data sets: IDRID, Kaggle and one local data set. Experiments are shown and indicate that RMCA U-net of this paper is superior to the other convolutional neural networks. The method in this paper is increased by 6% higher in PR-MAP than U-net on the IDRID dataset, increased by 10% in Recall than U-net on the Kaggle dataset and increased by 20% in F1-score than U-net on the local dataset. •An architecture with U-shape and residual module is proposed to segment exudates.•Multi-scale module is involved to learn the subtle feature from different layers.•Ultra-widefield fundus image dataset is introduced to validate the generalization.