Transfer Learning Approach for Classification of Diabetic Retinopathy using Fine-Tuned ResNet50 Deep Learning Model

Deep learning approaches have attracted a lot of attention as a way to classify retinal fundus images that include diabetic retinopathy (DR) because the old method of manual detection is labor-intensive and prone to misdiagnosis for large numbers of patients. The present AI medical models' inability to generalize when exposed to clinical data and the dearth of labelled medical data from which they can learn are also causes for concern. In a non-clinical setting, these methods have demonstrated good specificity and sensitivity for identifying Diabetic Retinopathy (DR). The task of determining the severity of diabetic retinopathy (DR) is studied in this research by fine-tuning the network to investigate the influence of transfer learning. Due of the numerous difficulties associated with medical annotation and privacy concerns, this study tests the automatic classification of diabetic retinopathy using the updated, fine-tuned ResNet50 model on the APTOS2019 dataset. Proposed Transfer learning approach outperforms existing methods in terms of Classification Accuracy, Precision, Recall, F1 Score. The proposed model achieves better results even with a smaller fraction of data, faster training and low computational resources, It's a strong point in favour of its novelty. In this study, the network is fine-tuned to examine the impact of transfer learning on the downstream task of assessing the severity of diabetic retinopathy (DR). The experimental results show that supervised pre-training on ImageNet followed by fine-tuning on labelled fundus images significantly boosts the efficacy of the medical image classifier when trained on full training data, demonstrating the effectiveness of transfer learning.