Experimental Comparison of State-of-the-Art Deep Learning Approaches for Oral Squamous Cell Carcinoma Cancer Segmentation

Oral Squamous Cell Carcinoma (OSCC) is a very common type of cancer that affects and can occur anywhere in the oral cavity. The segmentation of carcinoma cells from their non-carcinoma counterparts using deep learning architectures has been researched in some of the current studies, but further study is required in this area. In this research, the performance of the U-Net architecture and five modifications of the U-Net design is experimentally compared for the segmentation of OSCC images. High-quality Whole Slide Images (WSI) samples from the Oral Cancer (ORCA) dataset were used. Image augmentation and some other data preprocessing techniques were also used on the images before they were fed into the deep learning architectures. The architecture that gave the best performance is the U-Net with Inception-Resnet-v2 with an IOU score of 0.87, an F1 score of 0.86, a sensitivity of 0.86, and a specificity of 0.93. A key result of this research is that all of the U-Net modifications performed better when compared to the U- Net architecture.