Efficient Metric Learning with Graph Transformer for Accurate Colorectal Cancer Staging

Abstract

Colorectal cancer (CRC) is the third leading cause of cancer death in men and the third leading cause of cancer death in women in United States. So far, the histopathological image remains the golden standard in staging CRC, and accurate staging CRC is important for timely therapy and possible delay of the disease. Existing studies often utilized the pre-trained deep models to extract features from histopathological images, which neglected to take the supervised metric information into consideration. In addition, most of the existing methods did not take advantages of the correlations among different samples for the downstream classification tasks. To address the aforementioned problems, in this paper, we propose an efficient Metric learning with Graph Transformer (MGT), which adopts efficient metric learning to help extract distinguished image features followed by applying graph transformer for CRC staging. The main advantage of the proposed graph transformer is that it can fully exploit the correlations among different patients, which results in better tumor staging performance. To evaluate the effectiveness of the proposed method, we conduct several experiments for CRC staging on public available dataset TCGA-CRC in The Cancer Genome Atlas (TCGA). The experimental results show that our method can consistently achieve superior classification performance than the comparing methods.