Aiding Oral Squamous Cell Carcinoma diagnosis using Deep learning ConvMixer network

In recent years, Oral squamous cell carcinoma (OSCC) has become one of the worlds most prevalent cancers, and it is becoming more prevalent in many populations. The high incidence rate, late diagnosis, and inadequate treatment planning continue to be major concerns. Despite the enhancement in the applications of deep learning algorithms for the medical field, late diagnosis, and approaches toward precision medicine for OSCC patients remain a challenge. Due to a lack of datasets and trained models with low computational costs, the diagnosis, an important cornerstone, is still done manually by oncologists. Although Convolutional neural networks (CNNs) have become the dominant architecture for vision tasks for plenty of years, recent experiments show that Transformer-based models, most noticeably the Vision Transformer (ViT), may out-compete them in some settings. Therefore, in this research, a method called ConvMixer, which combines great features from CNNs and patches based on ViT was applied to an original very small dataset of only 1224 images in total for 2 classes, Normal epithelium of the oral cavity (Normal) and OSCC, 696 slides for 400x magnification and 528 slides for 100x magnification. However, the proposed models with small parameters and data augmentation performed magnificently, with 400x magnification (Accuracy: 99.81% - F1score: 99.87%) and 100x magnification (Accuracy: 99.62% - F1score: 99.77%).