Deep-Learning Cytology Tool Flags Malignant Oral Lesions With High Accuracy

Start Discussion
Posted: July 14, 2026

Deep-Learning Cytology Tool Flags Malignant Oral Lesions With High Accuracy

Edited by Dentaltown staff

A deep-learning system that analyzes oral brush cytology samples distinguished malignant lesions from healthy tissue with high accuracy in a study of 692 people, pointing to a potential minimally invasive way to help triage suspicious oral lesions. The findings were published July 13 in Scientific Reports.

The model generated a score called the oral cancer numerical index, which reached an area under the receiver operating characteristic curve of up to 0.99 when separating malignant samples from healthy controls. Reliability was high, with an intraclass correlation coefficient of at least 0.96 across repeated measurements.

Rather than relying on manual feature extraction and subjective review, the object-detection model directly classifies four cell types: differentiated squamous epithelial cells, small round cells, leukocytes, and lone nuclei. As disease severity increased, the proportion of differentiated squamous epithelial cells declined while small round cells and leukocytes rose, a pattern that tracked closely with histopathologic diagnoses (P < .0001).

The analysis drew on samples from 692 subjects spanning oral potentially malignant disorders such as leukoplakia and erythroplakia, oral squamous cell carcinoma, and healthy oral tissue. Conventional visual inspection alone is unreliable for judging which lesions carry the greatest risk, the authors noted, and cytology offers a less invasive alternative to biopsy for initial assessment.

The system is intended as an adjunct rather than a replacement for clinical examination and histopathologic diagnosis, and the authors called for prospective testing in routine dental and referral settings.

The study was led by Michael P. McRae of Custom DX Solutions and John T. McDevitt of the New York University College of Dentistry, with collaborators at the University of Texas and the University of Sheffield. Several authors, including McRae, McDevitt, and Spencer W. Redding, disclosed equity interests in OraLiva, a company developing oral cytology technology.

Sources:
Scientific Reports, “Deep learning single-cell analysis for cytologic evaluation of oral potentially malignant disorders,” July 13, 2026. DOI: 10.1038/s41598-026-47538-y. PMID: 42437755: nature.com/articles/s41598-026-47538-y


Deep-Learning Cytology Tool Flags Malignant Oral Lesions With High Accuracy

Views: 11
Start Discussion
Dentaltown Magazine
July 2026
Dentaltown Magazine July 2026 Read Current Issue Read Archived Issues
Sponsors
Townie Perks