Deep learning-based automatic segmentation of brain structures on MRI: A test-retest reproducibility analysis

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neuroscience

Background:

• Accurate segmentation of brain structures in MRI is crucial for diagnosing and monitoring neurological conditions.
• Deep learning models offer automated segmentation but their reproducibility across different MRI scanners needs evaluation.
• Variations in magnetic field strength (1.5T vs. 3T) may impact segmentation accuracy.

Purpose of the Study:

• To assess the reproducibility of deep learning-based automatic brain structure segmentation in MRI scans.
• To compare segmentation consistency across different scanner types and magnetic field strengths (1.5T and 3T).

Main Methods:

• Utilized deep learning techniques with human-in-the-loop for brain structure segmentation.
• Analyzed volumetric segmentation consistency across MRI scans from identical and different scanner types.
• Compared segmentation performance between 1.5T and 3T MRI scanners.

Main Results:

• High consistency in volumetric segmentation was observed when using the same scanner type (1.5T-1.5T or 3T-3T).
• Slightly better segmentation results were noted for 1.5T scanners compared to 3T scanners when used independently.
• Cross-comparisons between 1.5T and 3T scanners showed slightly reduced consistency, indicating magnetic field strength influence.

Conclusions:

• Deep learning-based segmentation demonstrates high repeatability when scanner type and protocol are consistent.
• Consistent scanner type and protocol are essential for reliable MRI studies, especially for brain atrophy monitoring.
• The findings support the efficacy of deep learning segmentation for clinical and research applications under standardized conditions.