Spine-GFlow: A hybrid learning framework for robust multi-tissue segmentation in lumbar MRI without manual annotation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Manual annotation for magnetic resonance image (MRI) segmentation is time-consuming, especially for multiple anatomical structures.
• Existing learning-based methods heavily rely on extensive manual annotations for supervision.
• Variations in image settings and pathologies complicate accurate segmentation.

Purpose of the Study:

• To develop Spine-GFlow, a hybrid framework for unsupervised multi-tissue segmentation of sagittal lumbar MRI.
• To combine Convolutional Neural Network (CNN) learned features with anatomical priors.
• To achieve robust segmentation independent of manual annotation and image variations.

Main Methods:

• A rule-based approach for automatic weak annotation (initial seed area) generation.
• A novel proposal generation integrating multi-scale image features and anatomical priors.
• A comprehensive CNN loss function optimizing pixel classification and feature distribution.

Main Results:

• Spine-GFlow was validated on two independent datasets (HKDDC and IVDM3Seg).
• The method achieved segmentation performance comparable to fully supervised models (mean Dice 0.914 vs 0.916).
• Accurate segmentation of vertebral bodies (VB), intervertebral discs (IVD), and spinal canal (SC) was demonstrated.

Conclusions:

• Spine-GFlow offers an effective unsupervised approach for lumbar MRI multi-tissue segmentation.
• The framework demonstrates robustness across different imaging settings and pathologies.
• This method significantly reduces the need for manual annotation in medical image segmentation tasks.