Leveraging retinal vessel segmentation for improved disease classification
Annotatsiya
This research study provides an accurate analysis of retinal disease detection on Fundus image vessel segmentation dataset (FIVEs) considering segmentation and classification task into the account. The development of an AI based application in ophthalmology can achieved by using FIVEs dataset due to it’s fined image resolution quality and precise annotations. ResUNet, U-Net, PSPNet and SegNet are the models being evaluated for the segmentation purpose. However, SegNet outperforms other models with efficient results of the highest mean intersection over union (Mean IoU) of 0.4558 and accuracy of 0.9335 and making it a best suitable model to perform this type of task. A custom designed convolutional neural network (CNN) outperformed other traditional models like VGG16, ResNet50, and EfficientNet with a classification accuracy of 91.6%, precision of 93%, recall rate of 0.88 and F1 score of 0.94. This demonstrates that the proposed model is most efficient in detecting and diagnosing different categories of retinal diseases. The novelty in this approach is the combined sophisticated methodology of segmentation and classification for a better accurate retinal diagnosis system. The performance of the classification model can be improved by the better features extracted from the segmentation task. This proposed study highlights the use of AI based solutions in the field of ophthalmic disease diagnosis and medical technologies can be considered for early treatment of detected disease. This proposed approach highlights the integration of both segmentation and classification task and achieves enhanced features extraction and gaining significant classification results as compared to previously proposed methodologies which concentrates on segmentation and classification task individually. • •Joint optimization of segmentation and classification • •SegNet achieved highest Mean IoU (0.4558) • •Custom CNN outperformed VGG16 and ResNet50 • •Achieved 0.94 F1-score in multiclass classification • •End-to-end intelligent diagnostic pipeline