Early diagnosis of oral cavity tumors using AI

The goal of an international study coordinated by the University of Sassari—and published in *Otolaryngology–Head and Neck Surgery*, the official journal of the American Academy of Otolaryngology–Head and Neck Surgery Foundation—is to use artificial intelligence to help physicians and dentists identify suspicious oral lesions early and more quickly refer patients for specialist evaluation. The study, titled "Image-Based Diagnosis of Oral Lesions: Performance of a Vision-Language Model versus Human Clinicians," evaluated the capabilities of a specially trained multimodal AI system based on Gemini 2.5 Pro in analyzing images of oral mucosal lesions.

This research is one of the outcomes of investments made by the University of Sassari and the Sassari University Hospital (AOU) as part of the "e.INS – Ecosystem of Innovation for Next Generation Sardinia" project—specifically within Spoke 01, "A new route to preventive medicine: genomics, digital innovation and telemedicine," for which the University of Sassari serves as the implementing entity and Spoke Leader. The program is funded under the National Recovery and Resilience Plan (PNRR) using European Union resources (NextGenerationEU). The Telemaco-S project—a collaboration between the University and the Sassari University Hospital dedicated to developing telemedicine and precision medicine solutions—also falls within this framework.

"Oral cavity carcinoma remains a significant problem because a still-excessive proportion of patients are diagnosed only when the disease is already at an advanced stage," explains Professor Giacomo De Riu, Full Professor of Maxillofacial Surgery at the University of Sassari and Director of the Maxillofacial Surgery Unit at the Sassari University Hospital (AOU). "Diagnostic delays can stem from the patient—who may underestimate an early-stage lesion—but also from the difficulty of recognizing symptoms that, in the initial phases, can resemble benign conditions. Reducing the time between the appearance of the lesion, its detection, and referral to a specialist is therefore one of the most important goals in the fight against these tumors."

The prospective multicenter study involved 350 patients recruited from 20 university and hospital centers across Italy, Belgium, France, Spain, and Israel. Lesion diagnoses were verified via histological examination, which served as the reference standard for evaluating the system's accuracy.

"We wanted to see what would happen when taking a publicly available multimodal model—specifically a version trained for this purpose—out of experimental datasets and into a real-world clinical setting," explains Luigi Angelo Vaira, a researcher at the University of Sassari and a medical specialist at the AOU of Sassari, who served as the study's principal investigator. "We developed a standardized protocol based on Gemini 2.5 Pro and tested it prospectively on 350 patients. We didn't simply ask the system to detect the presence of a lesion; it had to assess whether it was likely benign or malignant, formulate a diagnosis, and determine how urgently the patient needed to be referred for a specialist evaluation."

The results showed an overall accuracy of 97.1% in distinguishing between malignant and benign lesions.

The study also included an exploratory comparison with individual practitioners possessing varying levels of experience. In distinguishing between malignant and benign lesions, accuracy rates were 72.5% for general practitioners, 78.2% for dentists, 87.4% for oral pathology specialists, and 99.4% for expert head and neck surgeons, compared to 97.1% for the Gemini-based system.

"This comparison does not imply that artificial intelligence is superior to any specific professional group," Vaira emphasizes. "The study was not designed to demonstrate that. Rather, it highlights how complex the recognition of oral lesions can be outside of specialized settings. This is precisely where we envision a potential role for artificial intelligence: not to replace the physician, but to support them in identifying potentially suspicious lesions and facilitating faster referrals to specialists. The fact that the expert head and neck surgeon achieved the best results confirms that specialized expertise remains the gold standard."

The future outlook involves integrating tools of this kind—following further validation—into telemedicine and community-based triage pathways, thereby supporting the professionals who serve as the patient's first point of contact. Artificial intelligence is thus not proposed as a substitute for the physician, but as an additional tool capable of enhancing their capabilities, improving access to specialized expertise, and helping to reduce diagnostic delays.

"The key point," Vaira concludes, "is that we should not view artificial intelligence as a competitor to the physician. It is an extremely powerful technology that, if properly validated and incorporated into controlled clinical pathways, can become a valuable new tool at our disposal. The challenge for the coming years will be learning to use it correctly, placing it at the service of the professional and, above all, the patient."