Comprehensive benchmarking of deep learning architectures for multiclass histopathological classification of oral...

Accurate histopathological differentiation of normal oral epithelium, oral epithelial dysplasia (OED) and oral squamous cell carcinoma (OSCC) is essential but remains time consuming and prone to inter-observer variability. Deep learning provides a powerful approach for automated analysis of histopathological images in digital pathology. This study comparatively evaluated multiple deep learning architectures for automated classification of histopathological images into normal oral epithelium, OED and OSCC.

A curated dataset of 2363 histopathological images captured at 10x and 40× magnifications was analyzed, comprising normal epithelium (n = 1254), OED (n = 976) and OSCC (n = 102). The dataset was divided into training (80%) and testing (20%) sets. To address class imbalance and improve model robustness, data augmentation and weighted loss strategies were implemented, expanding the training dataset to 4750 images. Seven deep learning architectures- Vision Transformer (ViT), EfficientNet-B0, InceptionV3, ResNet-50, VGG16, MobileNetV3 and YOLOv11 were trained under standardized conditions using 224×224 input resolution for 100 epochs. Model performance was assessed using accuracy, precision, recall and F1-score.

EfficientNet-B0 achieved the highest validation accuracy (97.48%) and the best overall test performance with macro and weighted F1-scores of 0.97. InceptionV3, YOLOv11 and ResNet-50 also demonstrated strong classification capability (F1 ≈ 0.96). Misclassifications were rare and primarily occurred at histological transition regions between normal epithelium, dysplasia and early carcinoma.

Deep learning enables highly accurate classification of oral histopathological images. EfficientNet-B0 demonstrated superior performance, supporting the potential of AI-assisted digital pathology for early detection and objective assessment of oral precancer and cancer.