CMR_plan

This repository houses codes for automatic standard view planning for cardiac MR (CMR) imaging from a series of localizers via self-supervision by spatial relationship between views (see Fig. 1 below).

Note: we provide two versions of implementations, one using a vanilla UNet as backbone (main branch, corresponding to the MICCAI paper [1]) and the other using a stacked-hourglass backbone (SHG-Net branch, corresponding to the Med. Phys. paper [2]). Please be advised that you might need more training data for the SHG-Net to work effectively than for UNet.

Introduction

We provide an example data (case) located in the path "./data", which is used as both train and validation (test) data for code demonstration purpose.

Step 1: Preprocess localizers

The first step is to prepare the localizers based on which the standard views are prescribed, from raw DICOM files

cd prep_data
python prepare_2C_loc.py
python prepare_4C_loc.py
python prepare_SAX_loc.py
cd ..

The prepared data are saved in four folders named in the pattern "view_plan_data_xxx", in path "./data".

Step 2: Train regression networks

Next is to train view-specific regression networks for each localizer:

python train.py --view 2C
python train.py --view 4C
python train.py --view SAX

After training, the models are saved in "./models". We provide pretrained models as described in our paper, together with the code.

Step 3: Predict heatmaps with trained networks

Now, we can predict view-planning heatmaps for test data with the trained networks:

python inference.py --view 2C --prev models\2C_HT0.5.pth.tar
python inference.py --view 4C --prev models\4C_HT0.5.pth.tar
python inference.py --view SAX --prev models\SAX_HT0.5.pth.tar

The prediction results are saved in three folders named in the pattern "view_plan_pred_xxx_loc", in path "./data".

Step 4: Prescribe standard views by multi-view aggregation

The last step is to prescribe standard views by multi-view aggregation of the heatmap prediction results:

python prescribe_std_views.py

The prescribed planes are presented as ImagePositionPatient (IPP) and ImageNormalPatient (INP); please refer to the codes. For intuitive perception, we visualize the prescription results in the folder "prescribe_std_views" located in path "./data", where green color indicates the ground truth, red color indicates the automatic prescription, and yellow color indicates their overlap.

Acknowledgement

Kindly cite us if you use our codes in your projects:

1. Wei D, Ma K, Zheng Y. Training automatic view planner for cardiac MR imaging via self-supervision by spatial relationship between views[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer International Publishing, 2021: 526-536.
2. Wei D, Huang Y, Lu D, et al. Automatic view plane prescription for cardiac magnetic resonance imaging via supervision by spatial relationship between views[J]. Medical Physics, 2023, in print.