Abstract
Background: Artificial intelligence (AI) has great potential to detect fungal keratitis using in vivo confocal microscopy (IVCM) images, but its clinical value remains unclarified. A major limitation of its clinical utility is the lack of explainability and interpretability. Methods: An explainable AI (XAI) system based on Gradient-weighted Class Activation Mapping (Grad-CAM) and Guided Grad-CAM was established. In this randomized controlled trial, nine ophthalmologists (three expert ophthalmologists, three competent ophthalmologists, and three novice ophthalmologists) read images in each of the conditions: unassisted, AI-assisted, or XAI-assisted. In unassisted condition, only the original IVCM images were shown to the readers. AI assistance comprised a histogram of model prediction probability. For XAI assistance, explanatory maps were additionally shown. The accuracy, sensitivity, and specificity were calculated against an adjudicated reference standard. Moreover, the time spent was measured. Results: Both forms of algorithmic assistance increased the accuracy and sensitivity of competent and novice ophthalmologists significantly without reducing specificity. The improvement was more pronounced in XAI-assisted condition than that in AI-assisted condition. Time spent with XAI assistance was not significantly different from that without assistance. Conclusion: AI has shown great promise in improving the accuracy of ophthalmologists. The inexperienced readers are more likely to benefit from the XAI system. With better interpretability and explainability, XAI-assistance can boost ophthalmologist performance beyond what is achievable by the reader alone or with black-box AI assistance.
Introduction
Fungal keratitis (FK) is a leading cause of cornea-derived blindness globally. Rapid diagnosis is critical, yet traditional laboratory cultures often take days to yield results. In vivo confocal microscopy (IVCM) offers a real-time alternative, allowing clinicians to visualize fungal hyphae—the thread-like structures of the fungi—directly within the living eye.
Despite its potential, IVCM interpretation is notoriously difficult and requires years of specialized training. While deep learning models have demonstrated high accuracy in identifying FK, their clinical adoption is hindered by the "black box" problem: clinicians are hesitant to trust a probability score without understanding the underlying reasoning. This study investigates whether providing visual "explanations" via explainable AI (XAI) can improve clinical outcomes and doctor confidence.
Methods and Materials
The researchers developed a diagnostic pipeline using a ResNet-101 convolutional neural network. To provide interpretability, they implemented two specific XAI techniques:
- Grad-CAM: Generates a coarse heatmap highlighting the general regions of the image that most influenced the AI's decision.
- Guided Grad-CAM: Combines the heatmap with high-resolution pixel-level details to specifically highlight the thin, linear structures of fungal hyphae.
The study utilized a dataset of 3,177 IVCM images from the Guangxi Zhuang Autonomous Region People’s Hospital. The testing set consisted of 1,089 images, all adjudicated by a panel of three senior corneal specialists to ensure a rigorous ground truth.
Experimental Design
A randomized controlled trial was conducted with nine ophthalmologists categorized by experience: Experts (>15 years), Competent (>5 years), and Novices (3rd-year residents). They diagnosed images under three conditions:
- Unassisted: Original IVCM images only.
- AI-assisted: Images plus a probability histogram (the "black box" approach).
- XAI-assisted: Images, probability histogram, and the Grad-CAM/Guided Grad-CAM maps.
Key Results and Insights
Performance Gains
The standalone AI model achieved an impressive AUC of 0.983, with a sensitivity of 93.6% and specificity of 98.2%. When clinicians used the XAI system, the average diagnostic accuracy across all readers rose from 89.4% to 94.2%. The most dramatic improvements were seen in novice and competent readers; XAI assistance boosted novice accuracy to 90.7%, surpassing the performance of unassisted competent doctors.
Efficiency and Trust
While providing just a probability score (AI-assisted) increased the time doctors spent on an image, adding visual maps (XAI-assisted) reduced that time. The visual evidence allowed doctors to resolve the "why" behind a prediction faster than a raw number could. Furthermore, when the AI made a mistake on a negative image, the XAI maps often appeared "meaningless" or scattered. This visual incoherence helped doctors correctly identify the AI's error and maintain their own correct diagnosis.
Discussion and Limitations
The study highlights that while AI provides negligible benefits to experts who already perform at near-peak accuracy, it serves as a powerful "leveler" for less experienced clinicians. However, the authors acknowledge several limitations. This was a single-center study using a balanced dataset (1:1 positive to negative ratio), which may not reflect real-world clinical prevalence. Additionally, the study used static images, whereas clinical IVCM is a dynamic scanning process.
Future work will focus on multi-center validation and extending the XAI framework to other corneal pathogens, such as Acanthamoeba.
Conclusion
Explainable AI significantly improves the diagnostic accuracy of ophthalmologists identifying fungal keratitis. By "opening the black box," XAI provides the necessary interpretability to boost clinician performance beyond what is achievable by the reader alone or with standard black-box AI assistance.
Summary updated