NOTE: As of May 2024, this project is now archived in favor of this updated version that now includes model weights.

This repository provides code for training and running body composition estimation models on abdominal CT scans.

After cloning the repository, you have two options for setting up the environment. You may install all the necessary components directly on your system, or if you have docker on your machine, you may build a docker image that contains all the necessary requirements.

See the documentation pages for further details:

• Installation - For installing directly on your system
• Docker - For building and using the docker image
• Training - For training new models
• Inference - For running the model on new data

At this stage, we are not releasing our trained model weights publicly on github, and you will not find them in this repository. You are welcome to use this code on your own data to develop your own model, and you will find full instructions on how to do so in the documentation. We are however happy to discuss collaboration possibilities with investigators interested in using our models (including the deep learning models and population curve models) for their own studies. Please email us to discuss further:

• Chris Bridge, Massachusetts General Hospital (cbridge at partners dot org)
• Kirti Magudia, Duke University (kirti dot magudia at duke dot edu)
• Michael Rosenthal, Dana Farber Cancer Institute (Michael underscore Rosenthal at dfci dot harvard dot edu)
• Florian Fintelmann, Massachusetts General Hospital (fintelmann at mgh dot harvard dot edu)
• Camden Bay, Brigham and Women's Hospital (cpbay at bwh dot harvard dot edu)

This code accompanies the following publication:

Population-Scale CT-Based Body Composition Analysis Of a Large Outpatient Population Using Deep Learning To Derive Age, Sex, and Race-Specific Reference Curves

K. Magudia, C.P. Bridge, C.P. Bay, A. Babic, F.J. Fintelmann, F. Troschel, N. Miskin, W. Wrobel, L.K. Brais, K.P. Andriole, B.M. Wolpin, and M.H. Rosenthal

Radiology (In Press)

Article at RSNA

Furthermore, an earlier version of the same model was developed for the following publication:

Fully-Automated Analysis of Body Composition from CT in Cancer Patients Using Convolutional Neural Networks

C.P. Bridge, M. Rosenthal, B. Wright, G. Kotecha, F. Fintelmann, F. Troschel, N. Miskin, K. Desai, W. Wrobel, A. Babic, N. Khalaf, L. Brais, M. Welch, C. Zellers, N. Tenenholtz, M. Michalski, B. Wolpin, and K. Andriole

Workshop on Clinical Image-based Procedures, MICCAI, Granada 2018

Article at Springer Link, Article at Arvix

If you use this code in your publication, please cite these papers.

The Python code for body composition estimation was written by Christopher Bridge at MGH & BWH Center for Clinical Data Science. The z-score curve fitting R code in the stats directory was written by Camden Bay at Brigham and Women's Hospital.

The z-score fitting process associated with this work is available here.