Primary screening model for temporomandibular joint disc displacement based on triple-view face images

Temporomandibular joint (TMJ) disc displacement (DD) is a common clinical condition characterized by early onset and a high incidence rate. Its pathological changes impact facial contours, modify facial appearance features, and disrupt normal physiological activities of the face. Currently, MRI and CT are the most commonly utilized methods for TMJ examination, but unsuitable for primary DD screening in large sample populations. The utilization of deep learning techniques for primary screening and recognition of DD based on face images holds significant practical value. A total of 714 samples Triple-view Face-Image DD dataset (TvFID${}^3$) was established for this study, comprising 415 samples of disc displacement without reduction (DDwoR), 180 samples of disc displacement with reduction (DDwR), and 119 samples of normal TMJ (NOR). Each sample includes a series of triple-view face images encompassing frontal, left-side, and right-side views. Based on this facial DD dataset, the paper introduces a classification model for DD based on triple-view face images. All three-view face images undergo facial key-point and position recognition, followed by cropping and down-sizing, and finally face alignment. In this study, we focus on the automated discrimination of TMJ disc displacement disorders, specifically differentiating between DDwoR and DDwR subtypes. For model development, to identify DD cases within the general population, we trained a two-class (DDwoR+DDwR and NOR) classifier that achieved an accuracy of 86.3%. Subsequently, we deployed a secondary two-class (DDwoR and DDwR+NOR) classifier targeting DDwoR cases necessitating active clinical intervention in screening pipelines, which attained 90.4% classification accuracy. Our two-class (DDwoR and DDwR+NOR) discrimination system demonstrates strong potential for clinical application in primary screening of DD, offering reliable differentiation between pathological and normal cases as well as between different displacement subtypes. The DD discrimination model online (https://itmjtech.medi7.cn (DDwoR and DDwR+NOR) offers a cost-effective, efficient, and convenient solution for the primary screening of TMJ DD.