Abstract
Deep learning for glaucoma screening often relies on high-resolution clinical images and convolutional neural networks (CNNs). However, these methods face significant performance drops when applied to noisy, low-resolution images from portable devices. To address this, our work investigates ensemble methods using multiple Transformer architectures for automated glaucoma detection in challenging scenarios. We use the Brazil Glaucoma (BrG) and private D-Eye datasets to assess model robustness. These datasets include images typical of smartphone-coupled ophthalmoscopes, which are often noisy and variable in quality. Four Transformer models-Swin-Tiny, ViT-Base, MobileViT-Small, and DeiT-Base-were trained and evaluated both individually and in ensembles. We evaluated the results at both image and patient levels to reflect clinical practice. The results show that, although performance drops on lower-quality images, ensemble combinations and patient-level aggregation significantly improve accuracy and sensitivity. We achieved up to 85% accuracy and an 84.2% F1-score on the D-Eye dataset, with a notable reduction in false negatives. Grad-CAM attention maps confirmed that Transformers identify anatomical regions relevant to diagnosis. These findings reinforce the potential of Transformer ensembles as an accessible solution for early glaucoma detection in populations with limited access to specialized equipment.