Enhancing U-Net for Optic Cup and Disc Segmentation in Retinal Images Using Atrous Spatial Pyramid Pooling, Inception Modules, and Attention Gates

Image segmentation is essential in medical image analysis for glaucoma screening. Accurate delineation of the optic disc (OD) and optic cup (OC) in retinal fundus images is required for reliable clinical assessment. Manual segmentation is time-consuming and suffers from interobserver variability, which leads to inconsistent results. To address these limitations, this study proposes ASPP Inception Attention U-Net (ASPPIAU-Ne), an enhanced encoder-decoder architecture that integrates Atrous Spatial Pyramid Pooling (ASPP), Inception modules, and attention gates for feature selection in skip connections. The ASPP module is applied after the encoder to capture multi-scale contextual information and improve the representation of global and local structures. Attention gates suppress irrelevant features in skip connections and enhance important anatomical regions. In the decoder, Inception modules improve feature reconstruction and reduce upsampling artifacts. The proposed model is evaluated on DRISHTI-GS and REFUGE for optic disc, optic cup, and background segmentation. On the REFUGE dataset, the model achieves Dice scores of 88.8% for OD and 86% for OC with an IoU of 79.8% and 75.4%. On the DRISHTI-GS dataset, it achieves Dice scores of 80.2% for OD and 87.5% for OC with an IoU of 69.6% and 78.2%. For boundary evaluation, ASPPIAU-Net achieves Hausdorff Distance (HD) of 4.17 for OD and 3.94 for OC on REFUGE and 5.60 for OD and 5,61 for OC on DRISHTI-GS, indicating improved boundary alignment. For clinical consistency, the model achieves an MAE for VCDR of 0.005 on REFUGE and 0.086 on DRISHTI-GS. Overall, ASPPIAU-Net shows robust and balanced performance through multi-scale contextual learning and attention-based feature refinement. The model improves segmentation quality, particularly for the optic cup, and provides a reliable framework for automated glaucoma screening.