Smartphone Imaging System Enables Early Oral Cancer Detection

Oral cancer is often diagnosed at an advanced stage, reducing survival rates despite the mouth’s easy accessibility for visual exams. Many dentists and hygienists can identify abnormal lesions but lack the specialized expertise to distinguish benign from potentially malignant ones, leading to delays in referral and diagnosis. To address this challenge, researchers have developed a smartphone-based imaging system that combines autofluorescence and white light imaging with artificial intelligence (AI) to guide referral decisions.

The mDOC (mobile Detection of Oral Cancer) system, developed by researchers at Rice University (Houston, TX, USA), can assist dental providers in community clinics to accurately identify oral lesions that need further evaluation by cancer specialists. Designed for use in resource-limited settings, mDOC aims to improve early detection and treatment outcomes for oral cancer through a low-cost, portable, and easy-to-use solution.

The mDOC device integrates two imaging techniques—white light and autofluorescence imaging—into a compact smartphone attachment. Autofluorescence imaging employs blue light to visualize changes in tissue fluorescence, which can indicate abnormal or precancerous growth. Because benign conditions such as inflammation can also reduce fluorescence, the device uses a deep learning algorithm that combines imaging data with patient-specific factors like age, smoking history, and lesion location to improve diagnostic accuracy and generate referral recommendations.

In the study published in Biophotonics Discovery, researchers collected data from 50 patients at two community dental clinics. Each participant had up to five oral sites imaged with the mDOC device, resulting in hundreds of images for analysis. Expert clinicians reviewed the images and provided referral recommendations, which were used as ground truth for training and testing the AI model. The team employed a “rehearsal training” strategy, incorporating new data alongside previously collected images from high-prevalence and healthy populations to enhance real-world performance.

When tested on a low-prevalence dataset, the mDOC system correctly identified 60% of sites requiring specialist referral and avoided unnecessary referrals in most other cases. Notably, the AI model outperformed dental providers, who missed all cases that required referral. Two of the five lesions misclassified by the system had resolved before specialist evaluation, suggesting mDOC’s predictions may have been clinically accurate. However, the algorithm also generated 21 false positives, highlighting the need for refinement.

With an average imaging time of only 3.5 minutes per patient, mDOC integrates smoothly into routine dental practice. The researchers envision future iterations incorporating more comprehensive patient history and improved training data to further reduce false positives. The study underscores mDOC’s potential to serve as an accessible diagnostic support tool, particularly in clinics with limited access to oral cancer specialists, helping to facilitate earlier intervention and improved patient outcomes.

Related Links:
Rice University