Comparative Performance Analysis of Deep Learning Architectures for Oral Cancer Detection Using Histopathological Images

Abstract

Early and accurate detection of oral cancer is critical for improving patient outcomes. This study presents a comprehensive comparison of five convolutional neural network architectures—MobileNetV2, Xception, VGG19, ResNet50, and DenseNet201—for automated oral cancer detection from histopathological images. All models were fine-tuned through transfer learning using ImageNet pretrained weights and evaluated on 5,192 histopathological images from Kaggle. On this dataset, MobileNetV2 achieved highest classification accuracy $(\mathbf{9 0. 0 0 \%})$, followed by DenseNet201 (87.69%), VGG19 (86.54%), Xception (86.15%), and ResNet50 (83.27%). Statistical validation via paired t-tests confirmed significant performance differences $(p<0.05)$. Beyond accuracy, MobileNetV2 demonstrated 38% faster training, 84% smaller model size, and 33% faster inference, showing promise for resource-constrained clinical settings pending external validation.