Researchers have been using special maps called saliency maps to understand why certain AI models make specific decisions about glaucoma. They trained these AI models, called convolutional neural networks, to diagnose glaucoma using pictures of the eye's fundus, which is the inner lining of the eye. But instead of using full images, they simplified things by only showing the outlines of the optic disc and the cup within the eye. Surprisingly, these simpler images still helped the AI make accurate diagnoses, with scores ranging from 83. 31% to 88. 90% accuracy.

The study used 606 images and also looked at two other datasets called RIM-ONE DL and REFUGE. They tested nine different ways to create these saliency maps. To make the maps easier to understand, they used a special method to reduce noise and focus on key areas of the eye.

Interestingly, the maps varied quite a bit depending on the method and the model used. Sometimes, the areas highlighted by the AI didn't match the parts of the eye that doctors typically look at. However, overall, the results were pretty consistent, with strong correlations between the important areas in their dataset and the other datasets.

The findings show that while these saliency maps can help us understand why AI models make certain decisions, we need to be careful. They might be better for understanding general image relevance rather than making decisions about individual cases. Additionally, the regions the AI focuses on don't always match what doctors consider important for diagnosing glaucoma.

Simplified Eye Images Help Explain AI Glaucoma Diagnoses

Actions