Oral cancer is the most common type of head and neck cancer worldwide, leading to approximately 177,757 deaths every year. When identified at early stages, oral cancers can achieve survival rates of up to 75-90%. However, the majority of the cases are diagnosed at an advanced stage mainly due to the lack of public awareness about oral cancer signs and the delays in referrals to oral cancer specialists. As early detection and treatment remain to be the most effective measures in improving oral cancer outcomes, the development of vision-based adjunctive technologies that can detect oral potentially malignant disorders (OPMDs), which carry a risk of cancer development, present significant opportunities for the oral cancer screening process. In this study, we explored the potential applications of computer vision techniques in the oral cancer domain within the scope of photographic images and investigated the prospects of an automated system for detecting OPMD. Exploiting the advancements in deep learning, a two-stage model was proposed to detect oral lesions with a detector network and classify the detected region into three categories (benign, OPMD, carcinoma) with a second-stage classifier network. Our preliminary results demonstrate the feasibility of deep learning-based approaches for the automated detection and classification of oral lesions in real time. The proposed model offers great potential as a low-cost and non-invasive tool that can support screening processes and improve detection of OPMD.
Keywords: classification; convolutional neural network; deep learning; instance segmentation; leukoplakia; object detection; oral cancer; oral potentially malignant disorders; screening; semantic segmentation.