Generalizability of Deep Learning Architectures for Oral Squamous Cell Carcinoma Classification in Histopathology: A Comparative CNN-Vision Transformer Study with External Validation and Label-Quality Analysis.

Oral squamous cell carcinoma (OSCC) is the most prevalent oral malignancy, with delayed diagnosis contributing to poor survival. Deep learning has shown promise for histopathological detection of OSCC, but reported performance is frequently based on single-source datasets, with limited external validation and variable annotation quality. Our study was to compare representative CNN and transformer architectures for histopathological OSCC classification and assess their generalizability under external domain shift. A total of 10,817 H&E-stained image patches from 114 patients were classified as OSCC or normal/control. Images were independently reviewed by two oral pathologists, with discordant cases adjudicated by a senior pathologist (Cohen's κ = 0.975). Eight ImageNet-pre-trained architectures (ResNet18, ResNet50, ResNet101, DenseNet121, EfficientNet-B0, ConvNeXt-Tiny, ViT-B/16, and Swin-T) were fine-tuned using three random seeds. Models were trained, validated, and tested on the internally curated dataset {MIDAS} and subsequently evaluated on two independent external datasets used exclusively for external validation. Diagnostic performance, computational efficiency, statistical significance, and model interpretability were assessed. To investigate the observed reduction in external performance, cancer-labelled images from the external datasets underwent post hoc expert pathological review for potential label discordance. A total of 10,817 H&E-stained image patches from 114 patients were classified as OSCC or normal/control. Images were independently reviewed by two oral pathologists, with discordant cases adjudicated by a senior pathologist (Cohen's κ = 0.975). Eight ImageNet-pre-trained architectures (ResNet18, ResNet50, ResNet101, DenseNet121, EfficientNet-B0, ConvNeXt-Tiny, ViT-B/16, and Swin-T) were fine-tuned using three random seeds. Models were evaluated on a patient-level internal test set and two independent external datasets. Diagnostic performance, computational efficiency, statistical differences, and interpretability were assessed. Cancer-labelled images from External Dataset 2 were post hoc reassessed for label discordance. ConvNeXt-Tiny demonstrated the strongest overall performance. Substantial external degradation across models highlights the effects of domain shift and label quality, emphasizing expert annotation, independent validation, reproducible benchmarking, and multi-institutional datasets for clinical translation.