DEEP LEARNING-BASED CLASSIFICATION OF ORAL LESIONS USING CONVOLUTIONAL NEURAL NETWORKS: A COMPARATIVE STUDY OF CNN, RESNET50, AND CONVNEXT-SMALL ARCHITECTURES

Abstract

Background: oral mucosal lesions include various types of mucosal abnormalities , such as oral cancer (OCA), oral potentially malignant disorders (OPMDs) and benign lesions. Proper classification at early stage is essential for precise clinical management; however, conventional diagnosis depends on subjective visual examination. Objectives: To investigate and compare three deep learning models-a customized convolutional neural network (BasicCNN),ResNet50, and ConvNeXt-Small-for automated four-category classification of oral lesions from clinical images and to determine the impact of dataset augmentation, regularization, and statistical analysis on model performance. Methods: A dataset contains 1,304 clinically validated oral images curated from Piyarathne et al.(2024) along with the addition of 117 external OCA images from parola et al. (2024) to address the significant OCA underrepresentation (4.8%). One co-author collected 20 image as external validation dataset and performed an excisional biopsy for five cases.Three models were trained: BasicCNN (from scratch), ResNet50 (pretrained on ImageNet), and ConvNeXt-Small (pretrained, AdamW, CosineAnnealingLR, dropout, label smoothing ε=0.1, and 5-view test-time augmentation). Performance was assessed using macro F1, Cohen kappa, sensitivity, and specificity per class. Results: ConvNeXt-Small achieved the highest accuracy of 71.7% (95% CI: 65.9–77.0%), macro F1 of 0.719, and OCA F1 of 0.91 outperforming ResNet50 (69.8%, F1=0.697) and BasicCNN (53.2%, F1=0.538). OCA achieved AUC=0.950 (macro-AUC=0.881), with a sensitivity of 91.4% and a specificity of 96.9%. Pretrained models were statistically significantly better than BasicCNN (McNemar p < 0.001); the ConvNeXt vs. ResNet50 difference was not significant (p = 0.590). Dataset enrichment was essential: OCA F1 improved from 0.00 to 0.63 following enrichment. Conclusions: The ConvNeXt-Small architecture with progressive regularization and test-time augmentation outperformed other models in four-class oral lesion classification. Transfer learning played a vital role in boosting performance. The use of enriched data is crucial in order to detect OCA. These results can prove the possible application of deep learning in medical diagnostics.