matteo-dunnhofer / mrnet-baseline

Implementation of the MRNet model proposed in the paper "Deep-learning-assisted diagnosis for knee magnetic resonance imaging: Development and retrospective validation of MRNet" for knee disorders diagnosis.

MRNet replication

This repository reproduces (to some extent) the results proposed in the paper "Deep-learning-assisted diagnosis for knee magnetic resonance imaging: Development and retrospective validation of MRNet" by Bien, Rajpurkar et al.

Data

Data must be downloaded from MRNet's official website and put anywhere into your machine. Then, edit the train_mrnet.sh script file by expliciting the full path to MRNet-v1.0 directory into the DATA_PATH variable.

Execution

To perform an experiment just run

bash train_mrnet.sh

This will train three models for each view (sagittal, axial, coronal) of each task (acl tear recognition, meniscal tear recognition, abnormalities recognition), for a total of 9 models. After that, a logistic regression model is trained, for each task, to combine the predictions of the different view models. All checkpoints, training and validation logs, and results will be saved inside the experiment folder (it will be created if it doesn't exists).

Training and evaluation code is based on PyTorch and scikit-learn frameworks. Some parts are borrowed from https://github.com/ahmedbesbes/mrnet .