Laden: Lesion-Aware Adversarial Deep Network For The Grading Of Macular Diseases Using Color Fundus Images

Early detection of retinal illnesses such as diabetic macular edema (DME) and age-related macular degeneration (AMD) is essential for preventing central vision loss. Despite the fact that numerous solutions have been published, the performance of existing methods is inadequate due to the omission of lesion information in illness classification. A new lesion-aware adversarial deep network (LADeN) is proposed in this paper, which accurately exploits lesion information for clinically interpretable disease diagnosis. The first stage, which employs LADeN to efficiently reduce false positives, derives domain features from correct lesion segmentation maps. The extracted features from LADeN are combined with the baseline Efficient-Net-B7 model in the second stage to fine-tune disease classification and grading. Our proposed method exceeds the state-of-the-art in a range of tasks, including lesion segmentation, disease classification, and grading, as demonstrated by experiments on the IDRiD, MESSIDOR, and ADAM datasets.