Graph cuts with invariant object-interaction priors: application to intervertebral disc segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Computational Anatomy

Background:

• Accurate segmentation of intervertebral discs in lumbar spine MRI is crucial for diagnosing spinal conditions.
• Existing methods often require extensive training data or complex registration, limiting their clinical applicability.

Purpose of the Study:

• To develop and evaluate novel object-interaction priors for graph cut-based image segmentation.
• To apply these priors to the specific task of intervertebral disc delineation in lumbar spine MR images.
• To achieve translation, rotation, and scale invariance with efficient computation.

Main Methods:

• Optimization of a cost function incorporating learned geometric interaction priors.
• Development of translation, rotation, and scale-invariant priors based on distribution similarity and a fixed-point equation.
• Relaxation of an NP-hard problem formulation for efficient graph cut optimization.
• Validation using quantitative evaluations on 60 intervertebral discs from 10 subjects.

Main Results:

• The proposed algorithm achieved high correlation with expert manual segmentations.
• Demonstrated experimental invariance to translation, rotation, and scale.
• Required only a single subject for training, highlighting data efficiency.
• Achieved nearly real-time performance with few graph cuts.

Conclusions:

• The novel object-interaction priors are effective for intervertebral disc segmentation in MR lumbar spine images.
• The method offers robustness, efficiency, and reduced reliance on large training datasets.
• The demonstrated invariances support the algorithm's generalizability and minimal data requirements.