Exploring ResNet50 with Advanced Learning Strategies for Improved Ocular Disease Diagnosis

Diagnosing retinopathy accurately and promptly from fundus images is crucial, for preventing vision loss and determining the treatment. This study delves into how well deep learning models perform on the Ophthalmic Disease Recognition (ODIR) dataset in categorizing fundus images. Four models were analyzed namely a ResNet50 used as the baseline, ResNet with data augmentation, Bayesian Optimization and Learning Rate Scheduling for hyperparameter fine tuning. After testing the accuracy, sensitivity (recall) and specificity of these models were assessed to uncover their strengths and weaknesses. The Bayesian Optimization model stood out as the most effective achieving a 94% accuracy with excellent sensitivity and specificity. Data Augmentation also proved to enhance performance by enhancing accuracy and sensitivity. The performance boost from both models was statistically significant (at a = 0.05). This research highlights the importance of utilizing optimization methods like Bayesian Optimization to fine tune model hyperparameters for classification outcomes. These insights could have implications beyond retinopathy in medical image classification tasks paving the way for dependable models in clinical settings.