Segmentation Results: A Systematic Multi-Seed Evaluation of Preprocessing Strategies for Optic Disc and Cup Segmentation

Abstract

This archive contains the visual experimental results supporting the above publication. The study evaluates five preprocessing configurations for optic disc and cup segmentation using EfficientUNet++ with an EfficientNet-B7 encoder, trained and evaluated across three publicly available benchmark datasets: REFUGE, ORIGA, and Drishti-GS. Each configuration was trained independently under three random seeds (42, 15, 89) to characterise ranking stability. All quantitative results (Dice, IoU, Precision, Recall, mean ± SD across seeds) are reported in the main manuscript. This archive provides the supporting visual outputs, training curves, metric heatmaps, and prediction progression matrices, for each configuration and dataset. Note: Model weight files (.pth) are not included due to size constraints. They are available from the authors upon reasonable request. This archive does not redistribute the REFUGE, ORIGA, or Drishti-GS datasets. Only the experimental results generated in this study are provided. Users should download the original datasets from their official repositories and comply with their respective licenses.