Multiclass classification of oral mucosal lesions by deep learning from clinical images without performing any restrictions

Redondo, AlejandroIvaylova, KaterinaBachiller, MargaritaRincón, MarianoCuadra, Jose ManuelTamimi, FalehLópez-Cedrún, Jose LuisDiniz-Freitas, MárcioLago-Méndez, LucíaRubín-Roger, GuillermoTorres García Denche, JesúsBagan, LeticiaHernández Vallejo, GonzaloLópez-Pintor Muñoz, Rosa María2025-09-082025-09-082025-07-18Redondo A, Ivaylova K, Bachiller M, Rincón M, Cuadra JM, Tamimi F, López-Cedrún JL, Diniz-Freitas M, Lago-Méndez L, Rubín-Roger G, Torres J, Bagán L, Hernández G, López-Pintor RM. Multiclass classification of oral mucosal lesions by deep learning from clinical images without performing any restrictions. Biomedical Signal Processing and Control. 2026 En;111(108337)1746-809410.1016/j.bspc.2025.108337https://hdl.handle.net/20.500.14352/123750Oral cancer is a frequently malignant tumor that can be detected during an oral examination. Unfortunately, it is often diagnosed in advanced stages, which leads to low survival rates of about 50% at five years. Due to the low survival rate, it is crucial to develop automated systems that allow the classification of oral lesions according to their severity, aiding in the early diagnosis of oral cancer. This study aims to investigate the effectiveness of using clinical images and deep learning based models to perform a multiclass classification of oral mucosal lesions in color photographs taken without following any acquisition protocol. The classification differentiated four classes: malignant, potentially malignant, benign and healthy. The dataset included a total of 3246 images from 1013 patients, with 40 different categories of oral lesions, including healthy oral mucosa. The images showed different areas of the oral cavity and were captured from different perspectives by diverse dentists and maxillofacial surgeons in the practice. For the classification, different deep learning architectures were applied and compared, from the best known convolutional neural networks (CNN) and skip connection networks (SCN), to more innovative architectures such as visual transformers and a recent hybrid architecture, ConvNeXt v2. The ConvNeXt v2 Tiny architecture, with 85.53% accuracy, 85.02% precision, 85.50% recall, 84.92% F1-score, and 97.40% ROC AUC for an input image size of 354 × 354 pixels, outperformed the other architectures on the same database. The present model improved on previous proposals by considering a greater number of oral lesions and output classes.engAttribution-NonCommercial-NoDerivatives 4.0 Internationalhttp://creativecommons.org/licenses/by-nc-nd/4.0/Multiclass classification of oral mucosal lesions by deep learning from clinical images without performing any restrictionsjournal article1746-8108https://doi.org/10.1016/j.bspc.2025.108337https://www.sciencedirect.com/science/article/pii/S1746809425008481#da1open access613/614:[[778:616.314]:616-006.04]Images classificationOral cancerOral potentially malignant disordersDeep learningConvolutional neural networkSkip connection networksVisual transformersConvNeXtOdontología (Odontología)Salud pública (Medicina)OncologíaDiagnóstico por imagen y medicina nuclear32 Ciencias Médicas3299 Otras Especialidades Médicas3212 Salud Publica3213.13 Ortodoncia-Estomatología3207.13 Oncología2209.90 Tratamiento Digital. Imágenes