Application and Validation of a Smartphone-based Deep Learning System for Oral Potentially Malignant Disorders and Oral Cancer Screening

Application and Validation of a Smartphone-based Deep Learning System for Oral Potentially Malignant Disorders (OPMD) and Oral Cancer Screening

The goal of this clinical trial is to learn if smartphone-based deep learning system works to accurately detect oral potentially malignant disorders and oral cancer in adults. It will also learn about if it is as effective as assessments conducted by dentists and non-certified health provider.

We expect that the deep learning system will have higher sensitivity in detecting oral potentially malignant disorders and oral cancer, where as the dentists and non-certified health providers will exhibit higher specificity in screening.

Participants will be grouped into three arms: deep learning system (arm A) or board-certified dentist with deep learning system (arm B) or non-certified health providers (general practitioners) with deep learning system (arm C).

Oral cancer risk factors, such as habits of smoking or having chewed betel nut or alcohol drinking, would be recorded by anonymous questionnaires.

Background:

Oral cancer remains one of the leading causes of cancer-related deaths in Taiwan and worldwide. Artificial intelligence has the potential to improve oral cancer screening, enabling early detection by addressing healthcare access issues with high-quality solutions.

Objective:

To validate the smartphone-based deep learning system's accuracy in detecting oral potentially malignant disorders (OPMD) and oral cancer, while also demonstrating it is as effective as assessments conducted by dentists and non-certified health providers.

Methods:

Design, Setting and Participants: An open, three-arm, randomized controlled trial will be done in a medical center in Northern Taiwan between Jan 2025 to Dec 2025. The trial will include subjects aged 18 years or older who visit the cancer screening center for all kinds of screening. Oral cancer risk factors, such as habits of smoking or having chewed betel nut or alcohol drinking, would be recorded by anonymous questionnaires.

Interventions: Eligible subjects would be randomized in a 1:1:1 ratio using a computer-generated randomization algorithm to deep learning system (arm A) or board-certified dentist with deep learning system (arm B) or non-certified health providers (general practitioners) with deep learning system (arm C). The deep learning system in arm B and C would only be used for subsequent comparison and would not assist manual interpretation.

Main Outcomes and Measures: The primary outcome is the sensitivity and specificity for the three referral grades (benign (green), potentially malignant (yellow), and malignant (red)) by the deep learning system, dentists and non-certified health providers. The area under the curve (AUC) for each receiver operating characteristic (ROC) curve will also be calculated. The secondary outcome is subjects' feedback of comfortability during exam and the time needed for assessment.

Anticipated Results:

We hypothesize that deep learning systems will have higher sensitivity in detecting OPMD and oral cancer, whereas dentists and general practitioners will exhibit higher specificity in screening. The results could assist us in enhancing the oral cancer screening promotion process.

Inclusion Criteria:

• Adult patients (age ≥18) visiting cancer screening center

Exclusion Criteria:

• Unable to cooperate to fully open mouth/ navigate tongue
• Unable to cooperate for the assessment