Deep learning-based retinal image segmentation and classification for diagnosis of retinopathy of prematurity (ROP)

This thesis proposes a two-stage, deep learning-based approach for the diagnosis of retinopathy of prematurity (ROP). In the first stage, VesSegNet, based on the Attention U-Net architecture, was developed and enhanced with attention mechanisms for precise segmentation of retinal vessel structures. The model was trained on the DRIVE dataset, utilizing a hybrid loss function to effectively learn different vessel structures. The obtained segmentation outputs enabled detailed representation of fine vessel structures and successfully generated vessel maps on the previously unsegmented FARFUM-ROP dataset, which were then transferred for use in the second stage. In the second stage, the segmented vessel masks and original ROP fundus images were evaluated together to classify ROP plus disease (no-plus, pre-plus, plus). For this purpose, an EfficientNet-based dual-branch fusion architecture was designed. Experimental results reveal that VesSegNet achieves an F1 score of 0.807 on the DRIVE test set, whereas FusionROPNet attains a macro F1 score of 0.8137 for ROP plus disease classification, demonstrating robust performance. Given the clinical importance of vascular morphology in ROP diagnosis, segmentation information was incorporated into the classification pipeline.