Computational pathology-based artificial intelligence platform for the identification of common oral potentially malignant disorders

Oral potentially malignant disorders (OPMDs) refer to a group of significant precursor lesions for oral squamous cell carcinoma (OSCC). Early differential diagnosis of its subtypes including oral leukoplakia (OLK), oral lichen planus (OLP) and oral submucous fibrosis (OSF) is crucial for timely prevention and management of OSCC. The diverse clinical features of aforementioned conditions render clinical diagnosis challenging. To facilitate accurate and efficient diagnosis, this study aimed to construct a digital pathology-based artificial intelligence (AI) platform based on a sample of 1080 cases. Four deep learning models–Twins-SVT, ResNet18, ResNet50, and Inception_v3–were employed for patch-level analysis. The AI platform was developed by fusing deep learning-derived predictive features with multiple machine learning algorithms. The AI platform developed in this study exhibited good diagnostic performance, with the diagnostic area under the receiver operating characteristic curve of 0.870 (95% CI: 0.827–0.913), 0.810 (95% CI: 0.751–0.870), and 0.833 (95% CI: 0.770–0.897) for OLK, OLP, and OSF, respectively. In addition, this study investigated the correlation between deep learning-derived AI features and pathological findings, demonstrating excellent interpretability of AI in pathology. In this study, a robust AI platform was developed for differential diagnosis of OLK, OLP and OSF. Our research supports the feasibility of using deep learning methods to classify common OPMDs based on whole slide images.