Supervised fine-tuned approach for automated detection of diabetic retinopathy

The factors that concern the current AI medical models are the lack of generalizing capability when they are subjected to clinical data and also the scarcity of labeled medical data from which they can learn. This paper studies the role of transfer learning by fine-tuning the network when different fractions of medical data are available at the downstream task of diabetic retinopathy (DR) severity detection. The experimental results signify that supervised pre-training on ImageNet, followed by fine-tuning on labeled domain-specific fundus images significantly improves the efficacy of the medical image classifier when trained on full training data thereby suggesting transfer learning works. But what is less known is how the fine-tuning performance is affected when subjected to different fractions of data and if the learning is label efficient. Hence, we investigate the performance of the model under different fractions of labeled data (20 %, 40 %, 60 %, and 80 % of the entire data) on DR classification task, the results suggest that supervised fine-tuning underperforms when model is trained under low data regime. The proposed model achieves test accuracy of 0.8010, AUC of 0.86, F1 score of 0.6477, and cohen kappa score of 0.7007 when trained on full training data but underperforms when subjected to low data regime. Thereby suggesting the limits of supervised learning when the model is trained using limited annotated data. Hence our work opens door to further research in achieving good performance at low data regimes.