The eye is inflamed, the cornea clouding. The doctor needs to quickly determine the culprit — bacteria or fungus. The treatment depends on it. A mistake can cost sight. But the classic lab answer takes days, and in remote areas it may never arrive.
Researchers from the US and India propose a different approach: three images from a slit lamp — a device found in nearly every eye clinic. One shot in blue light, another in diffuse light, and a third in ordinary white light. These three "views" of the same opacity carry different information, like three interpreters listening to the same speech but catching different nuances. The model, named TPMKA, learns to reconcile these nuances. It first trains on pairs of images from the same patient to understand what they share and what is specific to a particular lighting. Then it fine-tunes for its task and finally merges all three data streams into a single decision.
The result: 86% accuracy. It may seem modest, but in a real clinic it's serious support. Crucially, the authors tested the model honestly: when evaluated on images from another country and another hospital, accuracy dropped. That's expected: cameras, settings, even the patient's skin tone affect the image. But even after rigorous cross-validation with data balancing, TPMKA remained the best among all tested approaches. Three images, three seconds, and the doctor gets a hint on how to start treatment while the lab is still preparing the culture.
as written by the authors arXiv:2607.03740
Microbial keratitis requires rapid pathogen identification to guide treatment, but culture- and PCR-based diagnostics are slow and resource-intensive. We developed a triple-phase multimodal framework for bacterial-versus-fungal keratitis classification using slit-lamp photographs acquired under blue-light, sclerotic-scatter, and white-light illumination, together with clinical metadata. The model combines cross-modality contrastive learning, modality-specific fine-tuning, and feature-level multimodal ensemble learning for patient-level prediction. We evaluated the framework on a multicenter dataset of 1,645 patients and 17,158 images from India and the United States. The model achieved 85.84% accuracy, 84.46% average F1-score, and 0.885 AUC. Site-specific evaluation showed that pooled results were overly optimistic, whereas resampling- and balance-based re-evaluation provided a more realistic assessment of cross-site generalization. Under all settings, our framework remained the top-performing approach. The code is available at https://github.com/yqwang01/TPMKA and dataset access will be provided subject to University of Michigan data-sharing clearance.
Blue light in eye photography isn't a whim. It makes damaged corneal tissue fluoresce, and the glow differs between bacteria and fungi.