Deep learning single-cell analysis for cytologic evaluation of oral potentially malignant disorders

Abstract

Abstract Oral potentially malignant disorders (OPMDs) such as leukoplakia and erythroplakia may harbor dysplasia or progress to oral squamous cell carcinoma (OSCC), yet visual inspection alone is unreliable for risk assessment. Cytology offers a minimally invasive adjunct, but conventional approaches depend on manual feature extraction and subjective review. Herein we report a deep learning (DL) object detection model that directly classifies four cell phenotypes: differentiated squamous epithelial (DSE) cells, small round (SR) cells, leukocytes, and lone nuclei. The DL model produced cytology-derived parameters that correlated strongly with histopathologic diagnoses across 692 subjects with OPMDs, OSCC, and healthy controls, including declining DSE cell proportion and increasing SR cells and leukocytes with disease severity ( p  < 0.0001). The oral cancer numerical index (OCNI) achieved AUROC values up to 0.99 for malignant versus healthy lesions, with excellent reliability (intra-class correlation coefficient ≥ 0.96). This reproducible, minimally invasive test provides a robust platform for early detection and surveillance of OPMDs.