Automated segmentation of myopic atrophic lesions on ultra-widefield fundus images: effects of training-set size and color representation

Abstract

Abstract Purpose To develop and evaluate a deep-learning model for automated segmentation of myopic atrophic lesions on ultra-widefield (UWF) fundus images and to evaluate the effects of training-set size and color representation on segmentation performance. Methods This study included 500 UWF images from highly myopic eyes with manually delineated myopic atrophic lesions. Independent training subsets of 50, 100, 200, 300, 400, and 500 images were constructed using a patient-level sampling strategy. For each subset size and color-space configuration (RGB, CIELAB, and luminance-only), models were trained 10 times with different random seeds. Segmentation performance was evaluated using the Dice coefficient and intersection-over-union (IoU). The effects of training-set size and color space were assessed using linear mixed-effects models, and performance variability was evaluated using Levene’s test. Results Training-set size significantly influenced segmentation performance ( p < 0.001). RGB and CIELAB inputs performed comparably, whereas luminance-only input showed modestly lower Dice and IoU than RGB input. Model performance improved with increasing dataset size, with progressively smaller improvements beyond approximately 300–400 images. The U-Net++ model in the primary experiment trained on 500 images achieved a mean Dice coefficient of 0.860 ± 0.006 and a mean IoU of 0.754 ± 0.009, demonstrating stable and reproducible convergence across runs. Conclusions Deep-learning-based segmentation of myopic atrophic lesions on UWF images is feasible. Performance gains became progressively smaller beyond approximately 300–400 annotated images under the present experimental conditions. RGB and CIELAB inputs showed comparable performance, whereas removal of chromatic information in the luminance-only condition was associated with modestly lower performance. These findings may inform efficient dataset design for quantitative UWF imaging analysis in highly myopic eyes and provide a methodological foundation for longitudinal investigation of lesion enlargement, risk-factor analysis, and spatial progression of pathologic myopia.