Moving Beyond CT Body Composition Analysis: Using Style Transfer for Bringing CT-Based Fully-Automated Body
Johannes Haubold1, Olivia Barbara Pollok, Mathias Holtkamp
1From the Institute of Diagnostic and Interventional Radiology and Neuroradiology, University Hospital Essen, Essen, Germany (J.H., O.B.P., M.H., L.S., C.B., J.S., B.M.S., K.B., J.K., M.O., L.U., M.F., F.N., R.H.); Institute for Artificial Intelligence in Medicine, University Hospital Essen, Essen, Germany (J.H., O.B.P., M.H., L.S., C.S.S., C.B., J.S., K.B., J.K., Y.W., M.O., L.U., M.F., F.N., R.H.); Institute for Transfusion Medicine, University Hospital Essen, Essen, Germany (C.S.S.); Center of Sleep and Telemedicine, University Hospital Essen-Ruhrlandklinik, Essen, Germany (C.S.S.); Data Integration Center, Central IT Department, University Hospital Essen, Essen, Germany (Y.W.); Department of Computer Science, University of Applied Sciences and Arts Dortmund (FHDO), Dortmund, Germany (C.M.F.); and Institute for Medical Informatics, Biometry, and Epidemiology (IMIBE), University Hospital Essen, Essen, Germany (C.M.F.).
This study introduces a deep learning method for automated body composition analysis (BCA) using MRI T2-weighted sequences, achieving high accuracy in segmenting various body regions and parts.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiology
Background:
- Deep learning (DL) is increasingly used for automated body composition analysis (BCA).
- Current DL methods often favor computed tomography (CT) over magnetic resonance imaging (MRI).
- There is a need for automated BCA methods using readily available MRI sequences.
Purpose of the Study:
- To develop and evaluate a deep learning approach for automatic BCA using MR T2-weighted sequences.
- To compare the performance of 2D and 3D deep learning models for BCA.
- To provide precise and detailed body composition information from MRI.
Main Methods:
- Generated initial segmentations by mapping CT data to synthetic MR images using CycleGAN.
- Trained nnU-Net V2 models (2D and 3D) on real T2-weighted MRI sequences.
- Refined segmentations with human annotators and evaluated performance using Dice and Hausdorff Distance metrics.
Main Results:
- The 3D ensemble model achieved high Dice scores for body regions (e.g., muscle 0.968, subcutaneous fat 0.98).
- The 2D ensemble model achieved high Dice scores for body parts (e.g., arms 0.952, torso 0.99).
- Stable performance was observed across all classes for both 2D and 3D ensemble models.
Conclusions:
- The developed deep learning approach enables efficient and automated BCA from T2-weighted MRI.
- The method provides accurate and detailed body composition data.
- This facilitates clinical research and potentially patient assessment.
