Convolutional Neural Networks for Accurate and Efficient Oral Histopathology Classification

Abstract

Accurate discrimination between leukoplakia with dysplasia, leukoplakia without dysplasia, and oral squamous cell carcinoma (OSCC) is central to risk stratification in oral pathology. This study evaluates four convolutional neural network (CNN) backbones under a uniform transfer-learning protocol for three-class histopathological image classification from NDB-UFES dataset. All images were resized to 224 $\times$ 224 pixels and trained using PyTorch with pretrained models obtained through the timm library. Data augmentation followed the default timm configuration associated with the selected model pipeline, and all experiments were executed on an NVIDIA RTX 5090 GPU with 32 GB of VRAM. The dataset comprised 3,763 images split into training, validation, and independent test subsets. ResNet-50, DenseNet-121, EfficientNet-B3, and Xception were compared using accuracy, precision, recall, F1-score, parameter count, computational cost, latency, throughput, and AUROC. EfficientNet-B3 achieved the highest test accuracy (0.8979) and precision (0.8918), whereas DenseNet-121 achieved the highest AUROC (0.9677). Confusion-matrix analysis showed that most residual errors occurred between dysplastic and non-dysplastic leukoplakia, while OSCC was recognized with comparatively high stability. These findings indicate that CNN transfer learning can provide a strong baseline for oral histopathology classification, although clinically robust deployment will require external validation and further investigation of diagnostically borderline leukoplakia cases.