Effect of image augmentation and proportional on multiple fundus image modalities for glaucoma prediction

This study develops an end-to-end pipeline for glaucoma prediction that works across both Standard View and Ultra-Wide Field fundus modalities. Modality is identified via a simple JPEG file-size rule and then a brightness-guided optic-disc localizer produces a centered optic disc cropping; across 1,300 images (714 glaucoma, 586 non-glaucoma), cropping succeeded on 1,281 images (98.54%), with failures occurring only in Ultra-Wide Field.We trained 42 models with each architecture: AlexNet, ResNet152, and Vision Transformers (ViT-B/16 and ViT-B/32) under consistent training settings, resulting in a total of 168 models, while systematically varying crop sizes (100%, 60%, 40%, 35%, and 30%) and augmentation strategies.Cropping to one-third of the original side (30–40%) consistently improves performance (accuracy 0.830–0.843; AUC up to 0.926–0.927) versus the 100% baseline (AUC 0.879), while maintain specificity.Regarding augmentation, one augmentation function per image increases the specificity–sensitivity balance, overlay-multiple augmentation decreases sensitivity/ AUC, and a 1:1 mix of originals with multiple augmentation achieves the best overall stability and scores (Acc 0.836, Sens 0.855, Spec 0.833, AUC 0.919).Finally, architecture matters: ViT-B/16 achieves the highest AUC (0.932) and sensitivity (0.889), ViT-B/32 achieves the highest specificity (0.828), and AlexNet outperforms the much deeper ResNet152 on all headline metrics (AUC 0.927 vs 0.882; Acc 0.839 vs 0.794).Overall, a simple modality split, robust optic disc centered cropping at 35–40%, balanced augmentation, and ViT-B/16 provide an effective, modality independent glaucoma detection pipeline for screening applications.