Deep-learning based model fitting and Gleason score lesion prediction.
If you use this repository in your diffusion MRI work please refer to this citation:
Vanya Valindria, Saurabh Singh, Eleni Chiou, Eleftheria Panagiotaki, et al. "Non-invasive Gleason Score Classification with VERDICT-MRI," 29th Annual Meeting of ISMRM, 2021.
Full text is available in the document.
- For deep learning based model fitting, most of the codes are in MATLAB. You can use the example scan (INN-104-RWB) to run the scripts. Run the following codes in order:
make_training_dataset_DL.m
to generate synthetic data from diffusion models under their own biophysical ranges. Output is the 'database...mat'
train_MLP_fitting.py
to train using the generated synthetic data (.mat) using a simple 3-layer Multi Layer Perceptron (MLP). Once you have the trained model (.sav), you can apply it on patient data (raw DW-MRI data, dependable on protocol).
preprocessing.m
to preprocess the (registration and denoising/unring etc) raw patient DW-MRI scans
post_process_DL.m
to obtain the ROI data for input to MLP training
apply_MLP_fitting.py
applying trained MLP to the input data (after being pre and post-processed)
save_maps.m
to convert from regression prediction from MLP to parametric maps of difussion MRI model
- For Gleason score (GS) prediction
Install MONAI first -> https://monai.io/
Run:
GS_classification.py
We need all/some of the parametric maps obtained from Step 1 and extract the pre-defined lesion ROI, and the ground truth (Gleason score for each lesion). We classify the lesion to 5-point Gleason score using SE-ResNet50 in MONAI, as shown as in the paper above, it gives better accuracy than DenseNet and resNet.