Transformer-Based Joint Classification Network for Diabetic Retinopathy and Diabetic Macular Edema

Within the scope of diabetic complications, diabetic retinopathy (DR) and diabetic macular edema (DME) are common eye diseases that pose a significant threat to patients' vision and are the main causes of permanent blindness in working-age groups. In clinical practice, accurate grading of DR and DME is particularly critical, because it is related to whether medical professionals can provide targeted treatment plans and management strategies for these diseases. Although the current joint classification method achieves considerable accuracy in classifying DME, it falls short in accurately classifying DR. This imbalance hinders overall classification performance improvement. Additionally, most classification models do not fully recognize the intrinsic link between DR and its complication, DME. This study presents an innovative cross-disease grading network that reveals intrinsic connections between different diseases and jointly evaluates DR and DME using only image-level supervision. The research uses a Transformer-based model to extract DR and DME features and fuses them using a self-attention mechanism module. This mechanism effectively captures and integrates complex features to improve classification accuracy. Experiments conducted on the public ISBI 2018 IDRiD challenge dataset verified our method, and the joint classification accuracy significantly improved to 63.14%. This result demonstrates that using Transformer for feature fusion is an effective strategy to optimize the joint classification performance of DR and DME.