AMD-Net: Automatic subretinal fluid and hemorrhage segmentation for wet age-related macular degeneration in ocular fundus images. (February 2023)

• Record Type: Journal Article
• Authors: Li, Pan; Liang, Lingling; Gao, Zhanheng; Wang, Xin
• Is Part Of: Biomedical signal processing and control. Volume 80(2023)Part 1
• Publication Date: 2023-02
• Subjects: Age-related macular degeneration -- Semantic segmentation -- U-Net -- Feature fusion -- Attention module; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104262
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: Age-related macular degeneration (AMD) is a chronic, progressively degenerative disorder of the macular. Wet AMD is responsible for 80 to 90 percent of all AMD-related blindness. Nowadays, ophthalmologists can diagnose wet AMD based on subretinal fluid and hemorrhage in ocular fundus images. However, this diagnosis process is time-consuming and influenced by subjective factors. Deep learning methods have achieved significant improvements for segmentation tasks in the retinal image, but only a few studies have focused on the subretinal fluid and hemorrhage lesion segmentation of wet AMD. This paper proposed AMD-Net, a novel U-Net architecture, which can segment the subretinal fluid and hemorrhage lesions of the wet AMD in ocular fundus images. The proposed AMD-Net consists of three components: encoder feature fusion unit (EFFU), skip connection block (SKB), and decoder attention block (DAB). The AMD-Net adopts the EFFU to extract and fuse the multi-scale features. And with the attention module, the EFFU can assign a higher weight for discriminative features. To reduce the semantic gap between the encoder features and decoder features, the AMD-Net designs an SKB. Furthermore, the AMD-Net introduces a DAB, a new module based on the UNet 3+ decoder, designed to leverage the local context information and enhance the contribution of high-level semantic features in tiny regions segmentation. We evaluate the proposed method on a private dataset collected by ourselves. Our model achieves 67.18% subretinal fluid Dice, 66.51% hemorrhage Dice, and 77.07% average Dice on this dataset. The experimental results indicate that our proposed AMD-Net is superior to the state-of-the-art deep learning methods. Highlights: AMD-Net is designed to segment subretinal fluid and hemorrhage lesions. Using the EFFU to improve the performance of small-scale lesions segmentation. A skip connection block is applied to capture the edge feature of lesions. A decoder attention block is used to extract relevant features in local information. The results indicated that AMD-Net performs best in lesions segmentation.