AI Framework for CMIL Grading
Published:Dec 27, 2025 17:37
•1 min read
•ArXiv
Analysis
This paper introduces INTERACT-CMIL, a multi-task deep learning framework for grading Conjunctival Melanocytic Intraepithelial Lesions (CMIL). The framework addresses the challenge of accurately grading CMIL, which is crucial for treatment and melanoma prediction, by jointly predicting five histopathological axes. The use of shared feature learning, combinatorial partial supervision, and an inter-dependence loss to enforce cross-task consistency is a key innovation. The paper's significance lies in its potential to improve the accuracy and consistency of CMIL diagnosis, offering a reproducible computational benchmark and a step towards standardized digital ocular pathology.
Key Takeaways
Reference
“INTERACT-CMIL achieves consistent improvements over CNN and foundation-model (FM) baselines, with relative macro F1 gains up to 55.1% (WHO4) and 25.0% (vertical spread).”