Accelerated muscle mass estimation from CT images through transfer learning

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Deep learning for medical image segmentation faces high data labeling costs.
• Computed tomography (CT) device variations hinder model generalizability.

Purpose of the Study:

• To develop an efficient deep learning strategy for medical image segmentation using CT scans.
• To address challenges in segmentation accuracy and model robustness across different CT devices.

Main Methods:

• Proposed transfer learning using SEED (small, manually labeled) images to train a VNet segmentation model.
• Evaluated VNet against UNet, UNETR, and Swin-UNETR models.
• Assessed model performance degradation due to catastrophic forgetting in multi-device training scenarios.

Main Results:

• Transfer learning effectively trained segmentation models with minimal data.
• VNet demonstrated superior performance in muscle and liver segmentation compared to other models and semi-automated tools.
• VNet exhibited the highest robustness against catastrophic forgetting.

Conclusions:

• CNN-based networks, particularly VNet, outperform transformer-based networks and semi-automatic tools for 2D CT image segmentation.
• The proposed strategy enhances efficiency and robustness in medical image segmentation.