Korean J Transplant.  2022 Nov;36(Supple 1):S92. 10.4285/ATW2022.F-2116.

AI model for the segmentation of skeletal muscle, visceral and subcutaneous fat at L3 level using donor CT

• Jang Y
• Jo EA
• Chun J
• Moon S
• Kim JS
• Han A
• Ha J
• Lee H
• Kim YC
• Min S

• KMID: 2535608
• DOI: http://doi.org/10.4285/ATW2022.F-2116

Abstract

Background
Although the L3 skeletal muscle index is accepted as a surrogate marker of sarcopenia and associated vulnerabil-ity, the effects of sarcopenia in kidney donors is not well defined. The purpose of this study was to develop and validate an auto- mated method to quantify the skeletal muscle, visceral and subcutaneous fat from an L3 slice on contrast-enhanced abdominal CT images of kidney donors.
Methods
The predonation arterial phase CT DICOM images of living kidney donors were downloaded and uploaded to 'OncoStudio' (OncoSoft Inc., Seoul, South Korea), which was used as the AI-based auto-segmentation tool. The AI model within the OncoStudio has a U-Net structure based on a 3D Dense block and automatically proceeds to CT site detection and segmentation without clicking by humans. For this study, a total of 41 datasets were used, 33 for training, one for validation, and seven for inde-pendent testing.
Results
The consistency between manually segmented volumes and automatically segmented volumes based on AI was evaluated. The average of dice similarity coefficient (DSC) representing the degree of agreement between 3D volumes was 0.92; skel-etal muscle: 0.92, subcutaneous fat: 0.97, visceral fat: 0.86. The average The Hausdorff distances 95% (HD95) representing the lower 95% distance between 3D surface points were 8.42, 3.81, and 1.72 mm, respectively.
Conclusions
An automated method for measuring volume of muscle and fats at L3 level was successfully developed. This auto-segmentation program can be easily used for prognostic evaluation including donor's sarcopenia and adult diseases.