Diabetic Retinopathy Detection and Grading: A Transfer Learning Approach Using Simultaneous Parameter Optimization and Feature-Weighted ECOC Ensemble

Early detection of Diabetic Retinopathy (DR) is crucial as it may cause blindness. Manual diagnosis of DR severity by ophthalmologists is challenging and time consuming. Therefore, there has been a significant focus on developing an automated system for identifying DR using retinal fundus images. Recent research has revealed that utilizing pre-trained deep learning networks for diverse image classification tasks provides notable benefits in this context. In this paper, a Transfer Learning (TL) approach with optimized feature weights and parameters is proposed for DR detection and grading tasks. To obtain better generalization during training and to optimize classification, features are extracted from the average pooling layers and fed to an Error Correction Output Code (ECOC) ensemble configuration. Two pre-trained networks (ShuffleNet and ResNet-18) are considered as each pre-trained network offers a different "point of view" of the fundus images, thereby providing more opportunities for accurate "grade-wise" discrimination. A simultaneous feature selection and parameter tuning of the ensemble is applied to further enhance the overall DR detection and grading. Adaptive Differential Evolution (ADE) is chosen for this purpose because it automatically configures the parameters, eliminating the need for manual parameter selection. In this paper, we evaluate two public domain datasets: 1) APTOS and 2) combination of EyePac + Messidor-2. Simulation results show that our proposed method performs better that the conventional deep learning models and are on a par with the existing research work. In particular, the optimal configuration for APTOS 5-class DR grading achieves an accuracy rate of 82%, while for APTOS 2-class grading, it achieves a higher accuracy rate of 96%. Finally, the best configuration for EyePac + Messidor-2 3-class grading results in 75% accuracy.