PiPNet3D: Patch-Based Intuitive Prototypes for Interpretable 3D Images Classification
We present PIPNet3D, a part-prototype neural network for volumetric images.
We applied PIPNet3D to the binary classification of Alzheimer's Disease from 3D structural Magnetic Resonance Imaging (sMRI, T1-MRI).
We assess the quality of prototypes under a systematic evaluation framework, propose new functionally grounded metrics to evaluate brain prototypes and develop an evaluation scheme to assess their coherency with domain experts.
Classes (clinical cognitive decline level):
- Cognitively Normal (CN)
- Alzheimer's Disease (AD)
arXiv preprint: "PIPNet3D: Interpretable Detection of Alzheimer in MRI Scans"
Accepted at the iMIMIC workshop during the MICCAI-2024 event.
Images and labels (cognitive decline level) were collected from the Alzheimer's Disease Neuroimaging Initiative (ADNI) https://adni.loni.usc.edu (data publicity available under request) and preprocessed using data_preprocessing.py functions.
Brain atlas (CerebrA) downloaded from https://nist.mni.mcgill.ca/cerebra/.
Codes adapted from the original PIPNet
Training a PIPNet: main_train_pipnet.py
Test a trained PIPNet: main_test_pipnet.py
Link to the weights of trained PIPNet(s) available in "models" folder.