Transforming glaucoma diagnosis: transformers at the forefront

Although the Vision Transformer architecture has become widely accepted as the standard for image classification tasks, using it for object detection in computer vision poses significant challenges. This research aims to explore the potential of extending the Vision Transformer for object detection in medical imaging, specifically for glaucoma detection, and also includes an examination of the Detection Transformer for comparative analysis. The analysis involves assessing the cup-to-disc ratio and identifying signs of vertical thinning of the neuroretinal rim. A diagnostic threshold is proposed, flagging a cup-to-disc ratio exceeding 0.6 as a potential indicator of glaucoma. The experimental results demonstrate a remarkable 90.48% accuracy achieved by the pre-trained Detection Transformer, while the Vision Transformer exhibits competitive accuracy at 87.87%. Comparative evaluations leverage a previously untapped dataset from the Standardized Fundus Glaucoma Dataset available on Kaggle, providing valuable insights into automated glaucoma detection. The evaluation criteria and results are comprehensively validated by medical experts specializing in the field of glaucoma.