Color Modifications and its Effect on Deep Learning Approaches for OSCC Cancer Segmentation

Oral Squamous Cell Carcinoma (OSCC) is the most commonly occurring oral cancer globally. Consequently, some state-of-the-art deep learning architectures, namely U-Net, U-Net with ResNet50, and U-Net with Inception have been proposed for OSCC segmentation. However, most of these approaches are computationally very expensive and/or suffer from sub-optimal results. Hence, an efficient and optimal approach for the segmentation of OSCC still remains an open problem. Consequently, Color Modification Techniques (CMT) have recently been found to give good performance when used for the semantic segmentation of crop and weed plants in images acquired from farming robots. In this paper, the effect of the use of CMT for preprocessing OSCC images on the performance of U-Net, U-Net with ResNet50, and U-Net with Inception is investigated. Testing is conducted using the publicly available oral cancer dataset with a size of 200 image samples and the corresponding ground-truth data gotten from the Cancer Genome Atlas (TCGA) dataset. Evaluation is conducted by using the Mean Intersection-Over-Union (mIOU), specificity, and sensitivity as metrics. An important contribution of this research is determining how the use of color modifications for preprocessing OSCC images impacts the performance of some deep learning architectures.

Jan 10, 2022 1:00 PM — Jan 12, 2022 3:00 PM