Project description: Project is dedicated to interpretation of deep learning models in medical domain. The main goal was to experiment with recent work on interpretable graph NNs -- IBGNN and check out it ability to extract biomarkers used for classification of Schizophrenia (COBRE dataset). Also, the experiments with non-learnable (baseline) methods of DL models interpretation were conducted on sMRI data (ADNI open-source dataset).
- Experiment 1: NNI AutoML for COBRE dataset.
- Experiment 2: Stability check with cross-validation (stratified, 7 folds) on COBRE dataset for 5 different random seed.
- Experiment 3: Repetition for cross-validation (stratified, 7 folds) on COBRE dataset.
Table of contents:
📈 Comet Training visualization
📃 Presentation
Future work:
- Finish analysis of the GNN results -- compare healthy/nonhealthy patient's brain networks, check on the consistency of network scores.
- Check GNN model behaviour on sMRI data
Current results:
Top ROIs for Schizophrenia illustrated above. On the left ROIs extracted by method, on the right some scientific foundings from here.