A Review of CAM-Based Visual Explanation on Diabetic Retinopathy

Abstract

Diabetic retinopathy (DR) is a leading cause of preventable blindness worldwide, and deep learning systems have achieved competitive performance for DR screening. However, the lack of transparent visual explanations remains a barrier to clinical adoption. Class Activation Map (CAM) methods are widely used to visualize model attention on fundus images, yet their behavior, evaluation, and clinical validity in DR remain fragmented across the literature. We conducted a PRISMA-ScR–guided scoping review of CAM-based visual explanation methods applied to DR classification and lesion-level analysis on color fundus photographs. We searched PubMed, IEEE Xplore, Scopus, Web of Science, and arXiv from 2016 to 2024, using combinations of keywords related to “diabetic retinopathy”, “class activation map”, and “explainable AI”. Studies were included if they applied CAM-family techniques to DR grading or lesion detection on retinal fundus images. The results are a total of 28 studies that met our criteria, covering 14 public and private datasets, predominantly EyePACS, Messidor, APTOS, IDRiD, and FGADR. We summarize the use of Grad-CAM, Grad-CAM++, Score-CAM, Ablation-CAM, LayerCAM, and several DR-specific CAM variants across common backbones (e.g., ResNet, EfficientNet, DenseNet) and categorize evaluation metrics into pixel- and lesion-level localization (IoU, Dice, DAUC/IAUC, ADD, DC/IC, ADCC) and causal faithfulness. We identify recurrent failure modes such as vessel-focused heatmaps, optic disc glare, and poor sensitivity to low-contrast microaneurysms. This review provides an evidence map of CAM-based explanations in DR, highlights gaps between visual plausibility and causal faithfulness, and proposes a compact “what works when” guideline relating DR task, dataset, backbone, and CAM choice. We present a clinical validity lens for lesion-level fidelity and referable DR triage, along with practical recommendations for standardized evaluation and clinically grounded CAM development in DR.