Background: Muscle-volume analysis is becoming essential for the planning of pre-surgical interventions, pre- and rehabilitation and recovery protocols (e.g., joint replacements), as well as for the design of training programs for performance enhancement or injury prevention. A tool to receive automated analysis of skeletal muscles after every regular MRI-scan is missing. Yet, there are numerous auto-segmentation deep learning models emerging specifically for medical images and muscles. However, in practice, these auto-generated segmentation labels require often fine-tuning by a human expert, which then helps to fine-tune or customize publicly available segmentation models.
Research questions:
Our project aims to enhance automated muscle segmentation in MRI scans by evaluating different deep learning models and weak-labeling strategies. Our main objectives include:
1. Model Comparison: We will compare pretrained models (e.g., Muscle-Map) with “trained-from-scratch” U-Net variants (e.g., U-Net, nnU-Net) in terms of accuracy, efficiency, and robustness for muscle tissue segmentation.
2. Weak-Label Analysis: We will investigate the effectiveness of various weak annotation strategies to improve segmentation models.
3. Model Improvement through Weak-Labels: We aim to assess how weak-labels can enhance segmentation quality.
