Flexible U-Net with Transfer Learning for Optic Disc Segmentation in Glaucoma Detection

Glaucoma, a progressive optic neuropathy leading to irreversible visual field defects and heightened intraocular pressure, underscores the critical need for early detection to prevent vision loss. Automating glaucoma detection can significantly reduce diagnosis time for ophthalmologists. Optic disc segmentation is an essential step in detection and diagnosis of glaucoma. In the experimental study, a pre-trained deep learning model called Flexible U-Net is utilized. This model incorporates EfficientNetB0 and EfficientNetB1 as the encoder components within the U-Net architecture. The dataset used for the evaluation of their efficacy is ORIGA. By analyzing model performance, it is found that EfficientNetB1 demonstrates superior accuracy in glaucoma detection compared to other models, emphasizing its potential for enhancing diagnostic efficiency in clinical practice. The results indicate that EfficientNetB1 U-Net has achieved impressive segmentation performance based on several metrics, including a dice coefficient of 0.9476, a Jaccard score of 0.9005, an accuracy of 0.9984, and an AUC of 0.99 on the ORIGA dataset.