Automated Deep Learning Framework for Retinal Landmark Detection and Quantitative Retinal Zoning in Ultra-Widefield...

Variability in ultra-widefield (UWF) fundus image acquisition and the lack of standardized post-acquisition processing complicate longitudinal monitoring and quantitative analysis because of differences in retinal coverage, image orientation, and acquisition-related artifacts. This study developed a deep learning-based post-acquisition framework for reproducible standardization and regional analysis of UWF images. We implemented a dual-model system consisting of a DeepLabV3Plus-based retinal region-of-interest (ROI) segmentation model to define the valid ROI and a U-Net-based heatmap regression model to localize the optic disc and fovea. The framework was developed using an internal dataset of 1381 UWF images and externally evaluated using 300 images from two publicly available UWF datasets, OUWFD and MSHF. ROI segmentation performance was evaluated using the Dice similarity coefficient (DSC), whereas optic disc and fovea localization accuracy was assessed using the normalized Euclidean distance (NED), defined as the centroid localization error normalized by the vertical optic disc diameter (DD). The detected landmarks were used as geometric anchors to automate rotation correction, define image-specific retinal zones, and generate an automated Early Treatment Diabetic Retinopathy Study (ETDRS)-inspired 7-field overlay. The ROI segmentation model achieved mean DSCs of 0.9938, 0.9702, and 0.9810 in the internal, OUWFD, and MSHF datasets, respectively. For optic disc localization, the mean NEDs were 0.0535, 0.0889, and 0.0795 DD, respectively, and all predictions were localized within 0.5 DD of the ground-truth centroid. For foveal localization, the corresponding mean NEDs were 0.1167, 0.1972, and 0.2136 DD, with 99.33%, 97.83%, and 95.62% of predictions localized within 0.5 DD, respectively. The pipeline corrected image orientation by a mean absolute rotation of 7.44° ± 4.78°. Zone 1 was fully contained within the valid retinal ROI in 97.33% and 96.67% of the internal and external test sets, respectively, whereas all seven ETDRS-inspired overlay fields were fully contained in 87.33% and 93.67%, respectively. The proposed framework integrates artifact-aware retinal coverage assessment, heatmap-based landmark localization, image orientation correction, and reproducible image-space overlays into a unified post-acquisition workflow. It provides a structured basis for quantitative retinal coverage assessment and reproducible image-space regional analysis of UWF images.