Code for IHD risk asessment from abdominopelvic CTs + EHR data, from the following paper (under review):
Opportunistic Assessment of Ischemic Heart Disease Risk Using Abdominopelvic Computed Tomography and Medical Record Data: A Multimodal Explainable Artificial Intelligence Approach
Authors: Juan M Zambrano Chaves, Andrew L Wentland, Arjun D Desai, Imon Banerjee, Gurkiran Kaur, Ramon Correa, Robert D Boutin, David J Maron, Fatima Rodriguez, Alexander T Sandhu, R Brooke Jeffrey, Daniel Rubin, Akshay S Chaudhari, Bhavik Patel.
Requires Linux. Create anaconda environment by running the following in command line:
conda create -n abct python=3.6.9
conda activate abct
while read requirement; do conda install --yes $requirement || pip install $requirement; done < requirements.txt
Install will take a few minutes.
To perform IHD risk assessment using an L3 dcm slice, clinical covariates or both:
-
Download the models from Google Drive locally.
-
Run the following script with appropriate flags, an example for fusion prediction is shown below:
python risk_assessment.py \
-risk_assessment both \
-trained_model_dir ./models/ \
-image_data_dir ./data/l3_slices \
-clinical_data_path ./data/clinical_data.csv \
-output_dir ./predictions/
Notes:
The dicom file name should be the anon_id specified in the clinical_data.csv file, i.e. abcde.dcm corresponds to anon_id : abcde.
If doing fusion inference, there should be a 1:1 correspondence between .dcm files and data rows in clinical_data.csv file.
2a. Sample data is provided with model weights. Running the example with the provided sample data will produce the following contents in prediction.csv (the inference will take a few seconds):
,I_C_fusion_1y,I_C_fusion_5y,clin_1y,clin_5y,img_1y,img_5y 0,0.003137208128035517,0.14864155105868518,0.009014944545924664,0.08169373869895935,0.03344954922795296,0.1995564103126526
Inference code been tested in Ubuntu 22.04.2 LTS with NVIDIA TITAN Xp GPU using CUDA 11.6.