The Clinical Value of Explainable Deep Learning for Diagnosing Fungal Keratitis Using in vivo Confocal Microscopy...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Artificial intelligence (AI) shows promise for detecting fungal keratitis from in vivo confocal microscopy (IVCM) images.
• Clinical utility of AI is limited by a lack of explainability and interpretability.
• Explainable AI (XAI) aims to address these limitations by providing insights into AI decision-making.

Purpose of the Study:

• To evaluate the clinical value of an explainable AI (XAI) system for fungal keratitis detection.
• To compare the performance of ophthalmologists with and without AI assistance, including explainable AI.

Main Methods:

• An XAI system using Gradient-weighted Class Activation Mapping (Grad-CAM) and Guided Grad-CAM was developed.
• A randomized controlled trial involved nine ophthalmologists (expert, competent, novice) assessing IVCM images under unassisted, AI-assisted, and XAI-assisted conditions.
• Performance metrics included accuracy, sensitivity, specificity, and time spent; AI assistance provided prediction probabilities, while XAI assistance added explanatory maps.

Main Results:

• Both AI and XAI assistance significantly improved accuracy and sensitivity for competent and novice ophthalmologists, without compromising specificity.
• The performance enhancement was more substantial with XAI assistance compared to standard AI assistance.
• There was no significant difference in the time required for image assessment between XAI-assisted and unassisted conditions.

Conclusions:

• AI demonstrates significant potential to enhance ophthalmologist accuracy in diagnosing fungal keratitis.
• Inexperienced ophthalmologists particularly benefit from XAI systems, improving their diagnostic capabilities.
• XAI's interpretability and explainability can augment ophthalmologist performance beyond unassisted or black-box AI-assisted assessments.