The prevention of cervical cancer relies on the accurate histopathological grading of its precursor lesions, particularly Squamous Intraepithelial Lesions (SIL), which can progress to Squamous Cell Carcinoma (SCC) if left undiagnosed or misclassified. However, the diagnostic evaluation of these lesions is often hindered by subjective interpretation, resulting in significant inter-observer variability. This inconsistency directly affects clinical decisions and emphasizes the urgent need for an objective, standardized diagnostic tool To address this issue, we developed an AI framework that mimics the diagnostic workflow of expert pathologists by integrating two core innovations. First, a domain-specific feature extractor, CerviScan-ViT, captures subtle patterns in the tissue architecture critical for distinguishing SIL from normal tissue and identifying high-risk lesions. Second, a reinforcement learning agent, SmartPatchRL, intelligently navigates whole slide images (WSIs) to perform a policy-driven search for diagnostically critical regions, enhancing both efficiency and focus. Validated in a multi-center, prospective and retrospective setting, our framework sets a new performance benchmark, achieving 90.6% accuracy during the testing phase and demonstrating robust generalization to external data. Notably, the model’s interpretable heatmaps highlight its ability to focus on true lesion-defining features while effectively disregarding slide preparation artifacts that often lead to over-grading in traditional models. This work offers a highly accurate, generalizable, and trustworthy solution poised to standardize SIL and SCC grading and improve the reliability of cervical cancer screening, ultimately aiding in early detection and prevention of both SIL and SCC.
Deng et al. (Thu,) studied this question.