Analysis of Tissue Heterogeneity in Graph Neural Networks

This project evaluates how graph construction methods affect a GNN's capability to capture spatial heterogeneity in cancerous tissue in the Grand Challenge on Breast Cancer Histology Images.

About the Dataset

The Grand Challenge included 400 microscopy images of 4 classes of breast cancer. For the purpose of this project, I used two classes (cancerous and non-cancerous). While the model achieved relatively high classification accuracy (around 90%), the primary focus of this work is on evaluating how different graph construction strategies preserve spatial heterogeneity patterns.

Repository Structure

• notebooks/

  • analysis.ipynb — Explanation of full experimental pipeline and visualization of results

• src/

  • graph_construction_helpers.py — Spatial graph construction methods
  • visualizers.py — All plotting and visualization utilities
  • data_helpers.py — Dataset loading and preprocessing
  • extract_features.py - Extracts and saves DenseNet-121 features from WSIs
  • macenko_normalizer.py - Helper class for macenko normalization
  • train_gnn.py — GNN model definition and training
  • test_heterogeneity.py - Analyzes tissue heterogeneity in DenseNet features and trained GNNs
  • test_gnn.py - Tests GNN ensemble accuracy

• data/

  • CSV outputs from heterogeneity experiments

• requirements.txt

How to Run

1. Install dependencies:

   pip install -r requirements.txt

2. Update the bach_directory path in constants.py to match the path to the Photos folder of the BACH dataset.
3. Open and run notebooks/analysis.ipynb. All figures and results will be generated.
4. The training pipeline from start to end involves running train_test_split.py → extract_features.py → train_gnn.py.
5. To rerun heterogeneity metrics, run test_heterogeneity.py.
6. To retrain models, run train_gnn.py.

The project also includes support for the BreakHis dataset, although no data from BreakHis was used in analysis due to variations in magnification between the two datasets. Update the breakhis_directory path in constants.py. Each file contains a flag that defines

dataset = "bach"

To run that file on BreakHis, change the flag to

dataset = "breakhis"

Methods Overview

• Patch-level feature extraction using pretrained CNNs
• Graph construction based on spatial and feature relationships:
  • Euclidean radius
  • Spatial k-NN
  • Delaunay triangulation
  • Feature k-NN
• Local heterogeneity quantified via neighborhood cosine dissimilarity
• GraphSAGE used to study learned aggregation effects

References

[1] Brussee, S., Buzzanca, G., Schrader, A. M. R., & Kers, J. (2025). Graph neural networks in histopathology: Emerging trends and future directions. Medical Image Analysis, 101, 103444. https://doi.org/10.1016/j.media.2024.103444

[2] Dolezal, J.M., Kochanny, S., Dyer, E. et al. Slideflow: deep learning for digital histopathology with real-time whole-slide visualization. BMC Bioinformatics 25, 134 (2024). https://doi.org/10.1186/s12859-024-05758-x

[3] Hameed, Z., Garcia-Zapirain, B., Aguirre, J.J. et al. Multiclass classification of breast cancer histopathology images using multilevel features of deep convolutional neural network. Sci Rep 12, 15600 (2022). https://doi.org/10.1038/s41598-022-19278-2

[4] Hamilton, William & Ying, Rex & Leskovec, Jure. (2017). Inductive Representation Learning on Large Graphs. 10.48550/arXiv.1706.02216.

[5] Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. (2017). Densely connected Convolutional Networks. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2261–2269. https://doi.org/10.1109/cvpr.2017.243

[6] Macenko, M., Niethammer, M., Marron, J. S., Borland, D., Woosley, J. T., Xiaojun Guan, Schmitt, C., & Thomas, N. E. (2009). A method for normalizing histology slides for quantitative analysis. 2009 IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 1107–1110. https://doi.org/10.1109/isbi.2009.5193250