BTS-Project A Combined Budding/T-Cell Score (BTS) in pT1 and Stage II Colorectal Cancer
  • HES-SO Fribourg
  • Linda Studer
  • Prof. Andreas Fischer
  • Geometric Deep Learning
  • Graph Neural Networks
  • Colorectal Cancer
  • Digital Pathology
Our skills
Deep Learning, Machine Learning, Graph-based Methods.
This project is funded by the Rising Tide Foundation.


RisingTide UniFr InstituteOfPathology
Example of a graph-representation of the lymphocytes and tumor buds in the tumor micro-environment.

Colorectal cancer (CRC) is the third most common malignancy in Switzerland and remains lethal in almost half of newly diagnosed patients. The most important parameter for prognosis and determining patient management is tumor spread according to the Tumor-Nodal- Metastasis (TNM) classification. However, the TNM system may show considerable differences in tumors within the same stage. This is particularly evident for stage II tumors, where 5-year survival ranges from 37%-67% 1. Therefore, additional biomarkers are needed for improved risk stratification in certain CRC patient subsets.

Tumor budding, the presence of single tumor cells and small clusters of tumor cells at the invasive tumor front, has been proposed to be such a biomarker. However, tumor buds are only part of the tumor micro-environment and do not account for other factors, such as the biological host response. Especially T-cell infiltrates have been extensively examined as a protective biomarker, leading to ‘Immunoscore’, a combined digital assessment of the density of CD3+/CD8+ T-cells in the center and the periphery of CRC. Our own preliminary data supports the above-mentioned study that a combined assessment of tumor budding and T-cell infiltrates as well-established biomarkers in CRC may be superior to either marker alone. So far the BTS (Budding-T-cell-Score) was calculated from each area as the average buds among all punches divided by the average of T-cell counts among all punches.

In this project, we want to further investigate the BTS score using graph-based deep learning techniques. In a first step, we need to detect tumor buds and lymphocytes on PanCK-stained whole slide images (WSIs). Secondly, graph-based representations of the WSIs are created based on the tumor buds and T-cells. Graphs allow modeling parts of an object with nodes and relationships between the nodes with edges. In our case, graphs will be used to model the structure of cells or cell regions. We believe that the graph-based representation will be beneficial for the combined assessment of tumor buds and T-cells, because it captures not only the raw count of the cells but also their structural arrangement, which, we hypothesize, might play a central role for predicting clinical endpoints. In a third step, geometric deep learning (see also, which extends deep learning beyond the domain of Euclidean geometry, is applied to the graphs to predict clinical endpoints, such as survival time.


  • Linda Studer, Shushan Toneyan, Inti Zlobec, Heather Dawson, Andreas Fischer, “Graph-based Classification of Intestinal Glands in Colorectal Cancer Tissue Images,” accepted at the MICCAI 2019 Computational Pathology Workshop COMPAY[Website] [PDF] [Bibtex]

    		title={Graph-based Classification of Intestinal Glands in Colorectal Cancer Tissue Images},
    		 author={Studer, Linda and Toneyan, Shushan and Zlobec, Inti and Dawson, Heather and Fischer, Andreas},


  • pT1 Gland Graph Dataset on GitHub