A hybrid evolutionary weighted ensemble of deep transfer learning models for retinal vessel segmentation and diabetic retinopathy detection

Segmentation of retinal blood vessels in fundus images is critical for early detection and treatment of diabetic retinopathy(DR). Due to the complex distribution of blood vessels, variations in noise, illumination, and vessel orientation in fundus images, the segmentation process becomes extremely challenging and time-consuming. In recent years, deep learning(DL)-based methods have been recognized as promising strategies for automatic retinal blood vessel segmentation that aids in early treatment, but most earlier DL algorithms prioritized accuracy over the complexity of the model for segmenting retinal vessels, making them challenging to adapt to medical devices. To the best of our knowledge, this is the first work to present a novel hybrid evolutionary weighted ensemble of three deep transfer learning(DTL) models: ResNet34, Inception V3, and VGG16 utilizing two public retinal fundus databases: DRIVE, and HRF for retinal blood vessel segmentation and latter detecting DR using Resnet34+Unet. Among comparing single models, Resnet34+U-Net achieved impressive accuracy of 0.9575, F1 and IOU scores of 0.8421 and 0.8159, and AUC values of 0.9868 in the DRIVE dataset. In comparison to the state-of-the-art results for segmenting retinal blood vessels using DRIVE and HRF datasets, our proposed approach achieves 0.9853 and 0.9816 accuracy, 0.9881 and 0.9833 precision, 0.9853 and 0.9828 recall, 0.8791 and 0.8497 F1 scores, 0.8066 and 0.7754 IOU scores, and AUC of 0.9996 and 0.9928 respectively, with a rise of nearly 7-8 % and for classification, Resnet34+Unet achieves an accuracy= 0.996, precision= 0.993, recall= 1, and AUC= 0.999 respectively. Hence, the proposed model shows good potential for real-time diagnosis.