Automated major psoas muscle volumetry in computed tomography using machine learning algorithms

Area of Science:

• Medical imaging
• Radiology
• Computational anatomy

Background:

• Psoas major muscle (PMM) volume is a valuable imaging biomarker in cross-sectional datasets.
• Manual segmentation for PMM volume calculation is time-consuming and labor-intensive.
• Automated methods are needed to efficiently assess PMM volume.

Purpose of the Study:

• To evaluate the accuracy of automated psoas major muscle volume (PMMV) quantification using two distinct methods: a generative adversarial network (GAN) and multi-atlas segmentation (MAS).
• To assess the performance of a combined GAN and MAS approach (COM) compared to individual methods.
• To determine the feasibility of automated PMMV measurement in CT datasets.

Main Methods:

• Manual segmentation of the bilateral psoas major muscle (PMM) by a radiologist in 34 abdominal CT scans served as the ground truth.
• Three automated methods were tested: a GAN-based approach (CNN), a MAS-based approach, and a combined approach (COM).
• PMM volume (PMMV) was calculated for each method, and results were compared to the ground truth using Dice similarity coefficient (DSC) and Spearman's correlation coefficient.

Main Results:

• The combined approach (COM) demonstrated highly accurate PMMV prediction (+0.7% deviation, p=0.33), outperforming individual GAN and MAS methods which significantly overestimated PMMV (+28.9% and +28.0%, respectively).
• The COM achieved a higher Dice similarity coefficient (DSC) of 0.82 [95%CI: 0.65-0.90] and Spearman's correlation of 0.73 compared to GAN (DSC: 0.75, correlation: 0.38) and MAS (DSC: 0.73, correlation: 0.62).
• Despite a small training dataset, the COM's segmentation accuracy approached that of manual segmentation.

Conclusions:

• The combined GAN and MAS approach effectively leverages the strengths of both individual methods for superior PMMV quantification accuracy.
• This hybrid segmentation strategy provides a significant improvement over isolated GAN or MAS implementations.
• The developed automated method shows promise for accurate and efficient PMMV assessment in clinical practice.