AI-Assisted Screening of Oral Potentially Malignant Disorders Using Smartphone-Based Photographic Images

Area of Science:

• Oral Medicine
• Artificial Intelligence
• Digital Health

Background:

• Oral potentially malignant disorders (OPMDs) and oral cancer are increasing globally, particularly in low- and middle-income countries.
• Limited resources for screening in remote areas lead to late detection, increasing mortality and reducing quality of life.

Purpose of the Study:

• To assess the effectiveness of AI techniques for OPMD detection using smartphone-captured oral cavity images in the Indian population.
• To evaluate deep learning models, including DenseNets and Swin Transformers, for classifying oral lesions.

Main Methods:

• A dataset of 2178 oral cavity images (1120 suspicious, 1058 non-suspicious) taken by trained healthcare workers was used.
• Deep learning models (DenseNet201, Swin Transformer) were trained and evaluated on an internal test set.
• The best model was further tested on an independent set of 440 images taken by untrained workers, assessing performance variations due to image quality.

Main Results:

• DenseNet201 and Swin Transformer (base) achieved high F1-scores of 0.84 and 0.83, respectively, on the internal test set.
• Performance decreased on the independent test set (images by untrained workers), with DenseNet201 achieving the best F1-score of 0.73.
• The AI model demonstrated potential in identifying suspicious oral lesions from photographic images.

Conclusions:

• AI-based analysis of smartphone images shows promise for OPMD screening in resource-limited settings.
• This simplified, image-based AI solution can support early detection and facilitate timely referrals for OPMDs.
• Further validation is needed, especially considering variations in image quality captured by untrained personnel.