Systematic application of saliency maps to explain the decisions of convolutional neural networks for glaucoma...

This paper systematically evaluates saliency methods as explainability tools for convolutional neural networks trained to diagnose glaucoma using simplified eye fundus images that contain only disc and cup outlines. These simplified images, a methodological novelty, were used to relate features highlighted in the saliency maps to the geometrical clues that experts consider in glaucoma diagnosis. Despite their simplicity, these images retained sufficient information for accurate classification, with balanced accuracies ranging from 0.8331 to 0.8890, compared to 0.8090 to 0.9203 for networks trained on the original images. The study used a dataset of 606 images, along with RIM-ONE DL and REFUGE datasets, and explored nine saliency methods. A discretization algorithm was applied to reduce noise and compute normalized attribution values for standard eye fundus sectors. Consistent with other medical imaging studies, significant variability was found in the attribution maps, influenced by the method, model, or architecture, and often deviating from typical sectors experts examine. However, globally, the results were relatively stable, with a strong correlation of 0.9289 (p < 0.001) between relevant sectors in our dataset and RIM-ONE DL, and 0.7806 (p < 0.001) for REFUGE. The findings suggest caution when using saliency methods in critical fields like medicine. These methods may be more suitable for broad image relevance interpretation rather than assessing individual cases, where results are highly sensitive to methodological choices. Moreover, the regions identified by the networks do not consistently align with established medical criteria for disease severity.