Performance Benchmarking of CNN Models in Oral Cancer Detection using Histopathological Images

Oral cancer is a malignant neoplasm that is both painful and challenging to diagnose in its early stages, making automated classification essential due to the underutilization of histopathologic images. This study addresses challenges like small datasets, overfitting, and computational expense by leveraging pre-trained deep learning models such as VGG19, MobileNetV2, and DenseNet121, which are effective in extracting meaningful features from large datasets like ImageNet while being computa tionally efficient. To enhance model generalizability and perfor mance, data augmentation techniques, including rotation, trans lation, shifting, shearing, zooming, horizontal flipping, brightness adjustments, additive random noise, and pixel normalization, are applied. Among the models, MobileNetV2 achieved the highest test accuracy of 84.13%, outperforming VGG19 (80.16%) and DenseNet121 (78.57%). The system demonstrated potential in providing reliable assistance to medical practitioners for early detection, with MobileNetV2 showing particular promise for computational efficiency and accurate classification. Future work aims to scale the approach for extensive and more varied datasets, enhancing its clinical applicability.