A robust deep learning model selection with data augmentation for automatic detection of tessellated fundus images and explainable artificial intelligence based interpretation

A robust deep learning system for automatically classifying retinal fundus images into two classes—normal and tessellated—is presented in this study. Visual Geometry Group – 16 is used as the base model, taking advantage of transfer learning to develop an efficient framework for fundus image classification. The approach uses nine different model architectures and makes use of a dataset of 352 fundus images that was increased to 4865 samples using sophisticated data augmentation techniques. Of these, Model_8 performed the best, achieving a loss of 0.129 % and an impressive accuracy of 99.39 %. The suggested approach ensures higher performance and dependability by combining rigorous data augmentation, efficient preprocessing, and model fine-tuning techniques. Furthermore, Explainable Artificial Intelligence was used to improve the interpretability of the model and visualize important aspects such as features or imposed pathologies in fundus images more clearly. The study offers promising support to ophthalmologists by offering precise automated diagnoses for the early identification and treatment of retinal disorders.