Graph Neural Networks for Colorectal Histopathological Image Classification

Representing data in array form is not always an efficient way. Sometimes it can cause loss of bias information that is inherent in data. With the help of taking an image to data matrix form as input instead of flattening to an array, deep learning methods, especially convolutional neural networks, are more successful than traditional machine learning techniques in terms of accuracy rate in image processing. Besides, graphs are a very efficient way to represent problems such as social networks, protein interface prediction, and images. Graph Neural Networks (GNNs) are deep learning methods that apply convolution logic to data like a graph, and many applications are efficiently applied. That is why representing histopathological images with graphs can be an advantage to know the connection between cores. The study uses GNNs to classify tissue types in the Chaoyang dataset. First, the superpixel graph is constructed from an image, and then GNNs models are applied to the constructed graph dataset. Experimental results present better accuracies than the compared literature methods.

I. Introduction

Cancer is one of the major threats affecting human life. The mortality rate in cancer diseases is quite high and its treatment is a very difficult process. Studies have shown how important early detection is in cancer treatment [1], [2]. The pathological examination is the gold standard for early cancer diagnosis and relies on pathologists' visual inspection to identify the tumor's origin, kind, and extent in the tissue. There are still a lot of obstacles to overcome, though. Pathologists assess the overall tissue as well as the structure, density, and diversity of the nuclei during the actual diagnostic procedure, which necessitates arduous workloads. Numerous variables, including pathologist tiredness and attention as well as the intricacy of the tissue structure, might have a detrimental impact on the accuracy of the diagnosis [3].