CONVOLUTIONAL NEURAL NETWORK-BASED MULTI-REGION SINGLE INSPECTION TECHNIQUE FOR AT-HOME, SELF-PRESCREENING OF ORAL/LARYNGEAL TUMORS

Abstract

Late detection of oral/laryngeal cancers or squamous cell carcinoma results in high patient mortality. Therefore, the detection of early-stage disease symptoms and timely medical treatment are important for improving long-term survival rates. Here, three deep learning models (single-shot detector, Yolo V4 and Tiny Yolo) were developed for target detection and binary type classification (normal/suspicious) for four representative oral/laryngeal regions (tongue, epiglottis, vocal cords, and tonsils) with a single-inspection process. The model performance was evaluated quantitatively on desktop and embedded platforms. We collected 1,632 endoscopic still-images and 20 diagnostic videos from the hospital database to train, validate, and test the models. Experimental results demonstrated that implemented models showed F1-scores ranging between 0.74–0.86, 0.86–1.00, and 0.74–0.87, and average precision ranging between 0.60–0.82, 0.92–1.00, and 0.72–0.98 for the tongue, epiglottis, and vocal cords, respectively, on the desktop platform. In addition, the Yolo V4 model showed performances of 0.92, 0.82, and 2.00 frames per second for the F1-score, average precision, and inference speed, respectively, on the embedded platform. Based on these results, we conclude that the implemented deep-learning-based at-home self-prescreening technique may be a reliable tool for personal oral/laryngeal healthcare, which will be especially important in endemic situations.