Abstract
There is a currently an unmet need for non-invasive methods to predict the risk of esophageal squamous cell carcinoma (ESCC). Previously, we found that specific soft palate morphologies are strongly associated with increased ESCC risk. However, there is currently no artificial intelligence (AI) system that utilizes oral images for ESCC risk assessment. Here, we evaluated three AI models and three fine-tuning approaches with regard to their ESCC predictive power. Our dataset contained 539 cases, which were subdivided into 221 high-risk cases (2491 images) and 318 non-high-risk cases (2524 images). We used 480 cases (4295 images) for the training dataset, and the rest for validation. The Bilinear convolutional neural network (CNN) model (especially when pre-trained on fractal images) demonstrated diagnostic precision that was comparable to or better than other models for distinguishing between high-risk and non-high-risk groups. In addition, when tested with a small number of images containing soft palate data, the model showed high precision: the best AUC model had 0.91 (sensitivity 0.86, specificity 0.79). This study presents a significant advance in the development of an AI-based non-invasive screening tool for the identification of high-risk ESCC patients. The approach may be particularly suitable for institutes with limited medical imaging resources.