A repository for the 02-740 Bioimage Informatics course project on GNN-based cancer detection and classification.
-
WIP: https://drive.google.com/file/d/18EfaqRfm8PwYN7Vg5b_r0sqqOF8FZt5o/view?usp=sharing
-
Pipeline:
- Images --> Node/Edge list --> Graphical Embedding (Pytorch Geometric/Node2Vec) -> GNN/Sk-learn Classifiers
-
We use RGB values on truncated portions of this dataset (https://www.kaggle.com/andrewmvd/malignant-lymphoma-classification) as node labels.
-
Euclidean pairwise distances are used as edge labels.
-
Graph is fully connected.
-
All of the above are good targets for improvement.
- Use a dataset with more heterogeneous classes with respect to nodes.
- Tweak the model archicture.