Dual-Branch Deep Learning Framework for Oral Cancer Detection from Lip and Tongue Images

Early detection of oral cancer plays a significant role in improving the patient's chances of survival. Visual screening of clinical photographs is a diagnostic tool that is currently being used; however, it is a difficult task given that there are different lighting conditions and the anatomical structures (and appearance of lesions) can be very different in lip and tongue pictures. All these difficulties affect the credibility of computer-aided diagnostic systems. In light of these circumstances, this work introduces a dual-branch deep learning architecture that exploits convolutional neural network features along with light texture descriptors for extracting global visual patterns and minute surface characteristics, respectively. The suggested method was tested on a public oral cancer image dataset consisting of lip and tongue photos. Results of the experiments show high classification effectiveness with an accuracy of 0.892, a macro-F1 score of 0.883, an AUROC of 0.912, and an AUPRC of 0.884. Probability calibration improved prediction confidence even more as it decreased the expected calibration error from 0.067 to 0.031. The results suggest that combining diverse visual features and adopting calibration techniques can significantly boost the performance of automated oral cancer screening. Through clinical decision-making, the proposed framework capable of being incorporated into cloud or mobile health systems and act as a tool for detecting oral cancer at its initial stage in telemedicine as well as community screening programs.