An Efficient Approach for Muscle Segmentation and 3D Reconstruction Using Keypoint Tracking in MRI Scans

arXiv Paper | Muscle Labeling Procedure

Magnetic resonance imaging (MRI) enables non-invasive, high-resolution analysis of muscle structures. However, automated segmentation remains limited by high computational costs, reliance on large training datasets, and reduced accuracy when segmenting smaller muscles.

Convolutional neural network (CNN)-based methods, while powerful, often suffer from:

• Substantial computational overhead
• Limited generalizability across diverse populations
• Poor interpretability

This study proposes a training-free segmentation approach using keypoint tracking with Lucas–Kanade optical flow, incorporating two keypoint selection methodologies: (1) manual selection based on visual observation and (2) semi-automatic selection using two-dimensional wavelet transform (2D DWT).

Performance

• Achieves a mean Dice similarity coefficient (DSC) ranging from 0.6 to 0.7, depending on the keypoint selection strategy
• Performs comparably to state-of-the-art CNN-based models
• Substantially reduces computational demands
• Enhances interpretability

This scalable framework presents a robust and explainable alternative for muscle segmentation in both clinical and research applications.

---

Usage

Dataset

• Data samples are available in the data folder.

Keypoints Selection Method

Method	Tracking Employing	Script
Manual Keypoint Selection	Initial selection	LKopflow_manual_select.py
	Reselection	LKopflow_manual_reselect.py
Semi-Automatic Keypoint Selection using 2D DWT	Initial selection	wavelet_tracking.py
	Reselection	wavelet_tracking_reselect.py