Convolutional Neural Network-Based Deep Learning Models for Diagnosis of Oral Squamous Cell Carcinoma and Oral Potentially Malignant Disorders: A Scoping Review

Background/Objectives: Oral squamous cell carcinoma (OSCC) represents the most common oral malignancy, accounting for over 90% of cases, with a five-year survival rate of approximately 50%. Oral potentially malignant disorders (OPMDs) carry an inherent risk of malignant transformation to OSCC, yet early and accurate diagnosis of both conditions remains challenging in clinical practice, owing to lesion heterogeneity and inter-observer variability in histopathological interpretation. While existing reviews have addressed artificial intelligence (AI) models’ diagnostic performance in oral oncology, these have typically combined varied model types and have not restricted inclusion to studies based on a held-out test set. This scoping review maps convolutional neural network (CNN)-based models applied with explicit diagnostic intent primarily to OSCC, with exploratory coverage of OPMDs, across histopathological and clinical photographic imaging, limited to studies reporting performance on held-out test data separate from model development. Methods: Following PRISMA-ScR guidelines, a systematic search was conducted in August 2025 across Scopus, PubMed, and the Cochrane Library. Studies were included if they applied CNN-based models with explicit diagnostic intent for OPMDs and/or OSCC using histopathological or clinical photographic images and reported quantitative diagnostic performance on a held-out test set. Extracted data included model architecture, imaging modality, dataset characteristics, and performance metrics comprising accuracy, sensitivity, specificity, precision, recall, F1-score, and AUC (area under the receiver operating characteristic curve). Results: Thirty studies published between 2020 and 2025 were included. CNN-based models reported generally favorable diagnostic performance, particularly with histopathological images, with accuracy frequently exceeding 90%. Models utilizing clinical photographic images demonstrated greater variability, with accuracy ranging from approximately 76% to 97.5%, in the seven of 13 studies that reported it. Across both modalities, binary classification tasks generally outperformed multi-class differentiation. Inconsistent metric reporting, reliance on small single-institution datasets, and near-absent reporting of external validation were observed as recurring methodological limitations. Conclusions: CNN-based models reported high diagnostic performance for OSCC, particularly in histopathological binary classification settings, although evidence for OPMD subtypes remains limited. However, translating this into clinical practice will require investment in large shared data repositories, standardized reporting of performance metrics, rigorous benchmarking against human experts, and external and prospective clinical validation.