Source-linked AI summary
CGC-Net: Cell Graph Convolutional Network for Grading of Colorectal Cancer Histology Images
Yanning Zhou, Simon Graham, Navid Alemi Koohbanani, Muhammad Shaban, Pheng-Ann Heng, Nasir Rajpoot
TL;DR
CRC grading requires glandular and tissue-microenvironment information that small image patches may not capture. CGC-Net converts large histology images into cell graphs, combines nuclear appearance, spatial interactions, adaptive multi-scale graph learning, and representative sampling, and reports state-of-the-art performance on a large CRC dataset.
Problem
Patch-based CRC grading methods may fail to incorporate the entire tissue micro-architecture needed alongside gland morphology and cell-level information.
Method
CGC-Net represents nuclei as graph nodes and cellular interactions as edges, using nuclear features, Adaptive GraphSage, and representative nuclei sampling.
Results
CGC-Net outperforms current state-of-the-art methods on a large-scale colorectal cancer grading dataset.
Takeaways & Limitations
Cell-graph modeling enables CGC-Net to aggregate cell morphology and tissue micro-architecture for colorectal cancer grading.
Abstract
from arXiv · showhide
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.
1. Introduction
CRC grading depends on gland formation and broader tissue context, but patch-based methods struggle to capture variable gland morphology and tissue micro-architecture. CGC-Net addresses this by representing nuclei and their interactions as a cell graph, with sampling and graph learning components, and reports superior performance on a large CRC dataset.
- Motivation: CRC grading distinguishes low- and high-grade adenocarcinoma by the degree of glandular formation and informs follow-up treatment and patient outcome assessment.
- Limitations of patch-based methods: Patch-based approaches trade image resolution against contextual coverage, while variable gland morphology makes a suitable patch size difficult to define.
- CGC-Net: Cell graphs represent nuclei as nodes and potential cellular interactions as edges, capturing cell-level information alongside tissue micro-architecture.
- Contributions: CGC-Net bridges deep learning with conventional cell graphs and uses nuclei rather than small patches to support biological interpretability.
- CGC-Net: CGC-Net uses nuclear appearance and spatial information, with a representative nuclei sampling strategy to reduce graph redundancy.
- Results: The proposed method outperforms other state-of-the-art methods in a comprehensive study on a large colorectal cancer dataset.
2. Related Work
Prior CRC grading methods use hand-crafted or deep-learning features, commonly classifying patches and aggregating their predictions. Context-aware and multi-scale methods expand the available context, while gland-based methods explicitly measure glandular morphology.
- Cancer grading in histology images: Earlier cancer-grading methods used hand-crafted nuclear appearance and morphological features, whereas recent methods widely apply deep learning across tissue types.
- Cancer grading in histology images: Typical deep-learning grading frameworks classify image patches with CNNs and combine individual patch predictions into an overall result.
- Contextual information: Multi-resolution, adaptive patch selection, and LSTM-based feature embedding have been proposed to incorporate larger or multi-scale contextual information.
- Colorectal cancer grading: For colorectal adenocarcinoma, BAM measures glandular morphology but relies on good-quality gland segmentation, while CA-CNN aggregates larger contextual regions using attention.
- Graph neural networks: GraphSage improves scalability through aggregation functions and batch training, while graph pooling methods reduce graph size for hierarchical graph classification.
3. Method
CGC-Net represents a large histology image as a cell graph whose nuclei provide node features and cellular interactions provide edges. Its architecture combines adaptive multi-scale graph convolutions, hierarchical graph clustering, and representative nucleus sampling for colorectal cancer grading.
- Cell graph construction: Nuclear instance segmentation and selected shape, appearance, and centroid features provide node descriptors for the cell graph.CIA-Net supplies nuclear masks, while feature selection retains 16 predictive nuclear descriptors before adding centroid coordinates.
- Cell graph construction: Representative nucleus sampling combines Farthest Point Sampling with random sampling to reduce graph redundancy while retaining nuclei from sparse areas.The experiments use a = 0.35 and b = 0.15, with b < a.
- Cell graph construction: Edges connect nuclei within a fixed Euclidean distance, with each node’s degree capped by its k-nearest neighbors.The adjacency rule requires both k-nearest-neighbor membership and distance below the threshold d.
- Graph representation and convolution: Graph convolution aggregates information from nuclei and their interactions in the nodes’ local neighborhoods.The method operates on non-Euclidean graph structure rather than a regular image grid.
- Graph network architecture: Graph clustering coarsens the graph into hierarchical nodes, and concatenated max-pooled representations from multiple stages feed a 3-class classifier.The network also applies reweighting to address over-smoothing and trains with cross-entropy loss.
- Adaptive GraphSage: Adaptive GraphSage uses stacked graph convolutions and attention-based weighting to fuse multi-level neighborhood features according to each node’s local structure.Intermediate representations are processed sequentially and combined through learned importance scores.
4. Experiment
Experiments evaluate CGC-Net on CRC histology data against state-of-the-art methods and through ablations of graph convolution, node features, and sampling. Results indicate that adaptive aggregation, combined nuclear features, and density-aware sampling improve graph representations and grading performance.
- Dataset and evaluation: The CRC dataset contains 139 images classified as normal, low grade, or high grade by gland differentiation.Images average 4548×7520 pixels at 20× magnification and are evaluated with three-fold cross-validation.
- Comparison with state-of-the-art: CGC-Net outperforms all competing methods by a large margin with smaller standard deviation on the CRC dataset.The comparison includes recent context-aware and domain-oriented approaches.
- Graph convolution: Adaptive GraphSage achieves the best patch-level and image-level accuracy among the compared graph convolution methods.It adaptively aggregates neighborhood information and combines multi-level features for irregular glandular structures.
- Nuclei sampling: Farthest sampling improves patch accuracy from 88.11% to 89.47% and image accuracy from 93.25% to 96.28% compared with random sampling.The proposed sampling strategy is designed to preserve tissue architecture while representing cell density and reducing graph redundancy.
- Cell graph visualisation: Hierarchical clustering aggregates local and global features, with learned clusters separating potential tumour and normal tissue regions.The visualisations assign nuclei in potential tumour regions to a red cluster and nuclei in normal regions to a green cluster.
5. Conclusion
The conclusion presents CGC-Net as a cell-graph approach for CRC histology grading that aggregates nuclear morphology and tissue micro-architecture. It combines graph convolution, clustering, and representative nuclei sampling, and outperforms current state-of-the-art methods on a large CRC dataset.
- Conclusion: CGC-Net aggregates cell morphology and tissue micro-architecture through graph convolution and clustering operations.The framework is presented as a first-of-its-kind cell graph convolutional network for colorectal cancer histology grading.
- Conclusion: Representative nuclei sampling reduces computational redundancy while preserving information about the tissue micro-environment.The method combines nuclear appearance features with local cellular interactions.
- Conclusion: CGC-Net outperforms current state-of-the-art methods on a large-scale colorectal cancer grading dataset.