CGC-Net: Cell Graph Convolutional Network for Grading of Colorectal Cancer Histology Images (1909.01068v1)

Abstract: Colorectal cancer (CRC) grading is typically carried out by assessing the degree of gland formation within histology images. To do this, it is important to consider the overall tissue micro-environment by assessing the cell-level information along with the morphology of the gland. However, current automated methods for CRC grading typically utilise small image patches and therefore fail to incorporate the entire tissue micro-architecture for grading purposes. To overcome the challenges of CRC grading, we present a novel cell-graph convolutional neural network (CGC-Net) that converts each large histology image into a graph, where each node is represented by a nucleus within the original image and cellular interactions are denoted as edges between these nodes according to node similarity. The CGC-Net utilises nuclear appearance features in addition to the spatial location of nodes to further boost the performance of the algorithm. To enable nodes to fuse multi-scale information, we introduce Adaptive GraphSage, which is a graph convolution technique that combines multi-level features in a data-driven way. Furthermore, to deal with redundancy in the graph, we propose a sampling technique that removes nodes in areas of dense nuclear activity. We show that modeling the image as a graph enables us to effectively consider a much larger image (around 16$\times$ larger) than traditional patch-based approaches and model the complex structure of the tissue micro-environment. We construct cell graphs with an average of over 3,000 nodes on a large CRC histology image dataset and report state-of-the-art results as compared to recent patch-based as well as contextual patch-based techniques, demonstrating the effectiveness of our method.

• The paper introduces CGC-Net, a novel framework that models histology images as cell graphs to enhance colorectal cancer grading by incorporating nuclear appearances and spatial features.
• CGC-Net converts large histology images into cell graphs using nuclear segmentation, Adaptive GraphSage for feature aggregation, and sampling, enabling comprehensive feature extraction on large images.
• CGC-Net achieves state-of-the-art results on colorectal cancer histology datasets, demonstrating superiority over traditional methods and holding potential for improving clinical reliability.

CGC-Net: Cell Graph Convolutional Network for Grading Colorectal Cancer Histology Images

Colorectal cancer grading typically assesses the degree of gland formation within histology images. Existing methods often focus on small image patches, failing to capture the complex tissue micro-environment necessary for accurate grading. This paper introduces a novel framework, the Cell Graph Convolutional Network (CGC-Net), which models histology images as graphs, enhancing cancer grading by incorporating both nuclear appearances and spatial features. CGC-Net converts each large histology image into a cell graph where nodes represent individual nuclei and edges denote cellular interactions assessed by node similarities.

CGC-Net provides a significant advancement in modeling the intricate tissue micro-environment by considering both local cellular interactions and global tissue architecture. This is achieved through several innovations, including nuclear segmentation to define graph nodes, Adaptive GraphSage for multi-scale feature aggregation, and a sampling technique to address redundancy by removing excess nodes in dense regions. The use of graph theory allows CGC-Net to process images substantially larger than those typically used in patch-based approaches, thereby enabling more comprehensive feature extraction.

The paper reports state-of-the-art results on a large colorectal cancer histology image dataset, demonstrating the superiority of CGC-Net over traditional patch-based methods. The network, equipped with Adaptive GraphSage, provides an adaptive fusion of node features, enhancing its capability to model various scales of tissue architecture. Experimental results indicate that CGC-Net achieves high accuracy in both patch-level and image-level evaluations.

The implications of this research are substantial for the field of computational pathology. By accurately modeling the interactions and features at a cellular level, CGC-Net can improve the reliability of cancer grading, potentially reducing inter-observer variability in clinical settings. This approach could be extended to other types of cancer, leveraging graph-based modeling for better diagnostic accuracy.

Future work could explore the application of CGC-Net to other histological patterns and cancer types, expanding its utility in the domain of digital pathology. Moreover, integrating deep learning frameworks with graph-based models could further refine the precision of tissue analysis, paving the way for advancements in automated cancer grading and improved patient outcomes.