AN EXPLAINABLE FRAMEWORK FOR DETECTION OF DIABETIC MACULAR EDEMA USING DEEP LEARNING TECHNIQUES

The proposed research work aims to provide a deep learning model for the early and accurate detection of diabetic macular edema (DME). This research work proposes the utilization of VGG19 along with the gradient-weighted class activation mapping (Grad-CAM) technique for the detection and detailed classification of macular edema. During the pre-processing stage, the Bilateral filtering model technique is used for contrast enhancement, and the generative adversarial network (GAN) is used for the generation of synthetic retinal images. The GAN technique is capable of augmenting limited datasets. For segmentation of the macula region in retinal pictures, the Neural Alpha Matting technique is utilized. This technique is mainly useful for refined boundary segmentation. Once the macular region is segmented, it is then fed into VGG19 for feature extraction and classification, as well as the Grad-CAM for the purpose of providing a comprehensive diagnosis of the presence of macular edema. The proposed work produces better experimental results, with 98.9% accuracy, specificity of 98.6%, sensitivity of 98.7%, precision of 98.6%, and 98.7% F1-score. The integration of GAN-based augmentation, Neural Alpha Matting segmentation, and Grad-CAM explainability provides both diagnostic accuracy and transparent decision support, aligning with the emerging paradigm of interpretable medical AI.