Single slice thigh CT muscle group segmentation with domain adaptation and self-training

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Thigh muscle segmentation is crucial for evaluating muscle anatomy, metabolic disorders, and aging.
• Manual annotation of muscles in MR images is labor-intensive and challenging to transfer to CT images.

Purpose of the Study:

• To develop an unsupervised domain adaptation pipeline for thigh muscle segmentation on single-slice CT images using MR image labels.
• To overcome the challenge of transferring annotations from 3D MR to 2D CT thigh images.

Main Methods:

• Utilized CycleGAN for MR to CT image appearance transformation.
• Implemented a self-training strategy with hard and easy pseudo-label cohorts for segmenter fine-tuning.
• Refined pseudo-labels based on anatomical assumptions.

Main Results:

• Achieved a mean Dice score of 0.888 across multiple thigh muscle groups (gracilis, hamstrings, quadriceps femoris, sartorius).
• Demonstrated robust performance on 152 withheld single CT thigh images.

Conclusions:

• This is the first pipeline for domain adaptation from MR to CT for thigh images.
• The proposed method effectively segments thigh muscles on 2D single-slice CT images.
• The pipeline is publicly available on GitHub.