Tumor budding with graph neural networks
Aim
Tumor budding and CD8+ lymphocytes, also known as cytotoxic T-cells, are essential factors in colorectal cancer. By studying them together, we can gain a complete understanding of the micro-environment and the biological response. Our aim is to use graph-based deep learning to investigate their spatial layout as a potential risk predictor for pT1 and stage II CRC patients. Graphs will be used to model the topology of the tumor buds and T-cells, providing a more comprehensive assessment that includes the cells' structural arrangement in addition to their raw count. To stratify patients, we will train Graph Neural Networks (GNNs) on the graph representations. For pT1 patients, we hope that this system will help pathologists make better-informed decisions on whether patients are high-risk and need colon resection, as these cancers are minimally invasive. For stage II patients, we have observed high variability in survival rates and aim to predict them better, as high-risk patients could benefit from adjuvant chemotherapy. By combining tumor budding and T-cell infiltration with graph-based deep learning, we hope to develop a more accurate risk stratification tool that can improve patient outcomes.
Methods
To capture the spatial layout of the tumor microenvironment, we built the pT1 Hotspot Tumor Budding T-cell Graph (pT1-HBTG) dataset (626 tumor budding hotspots from 575 pT1 CRC patients) collected from eight pathology institutes. On each slide, tumor buds and CD8+ T-cells were automatically detected, and an expert pathologist selected the budding hotspot at the invasive front. Every detected cell became a node in a graph, with edges connecting neighbouring cells. We compared three ways of building these edges (Delaunay, Delaunay-Star and Hierarchical) and different combinations of node information (cell type, position, and a ViT-16 image embedding). Graph Neural Networks (GraphSAGE, GIN and GATv2) were then trained to classify each patient as high- or low-risk, that is, whether, in retrospect, they needed the colon resection. All models were benchmarked against the current Swiss Society of Gastroenterology (SGG) risk guidelines. Both the dataset and the code are publicly available.
Results
The best-performing graph model increased specificity by around 20% compared to the current clinical guidelines, without any loss in sensitivity. In practice, this means correctly sparing many more low-risk patients from surgery while still catching nearly all of the high-risk ones. The spatial arrangement of tumor buds and T-cells therefore carries prognostic information beyond their raw counts, and a graph-based analysis could help reduce overtreatment in pT1 CRC.
| Method | Specificity | Sensitivity |
|---|---|---|
| SGG clinical guidelines (baseline) | 22% | 85% |
| Best graph model (GraphSAGE + Hierarchical graph) | 43% | 84% |
Members
Linda Studer
Inti Zlobec
Heather Dawson
Andreas Fischer
Rolf Ingold
Funding source
Publications
Studer L, Bokhorst J-M, Nagtegaal I, Zlobec I, Dawson H, Fischer A. Tumor Budding T-cell Graphs: Assessing the Need for Resection in pT1 Colorectal Cancer Patients. Medical Imaging with Deep Learning (MIDL), Proceedings of Machine Learning Research (PMLR), vol. 227, pp. 235–259, 2024. https://proceedings.mlr.press/v227/studer24a.html