Fast surface reconstruction of human brain MRI: benchmarking deep-learning based morphometry tools

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Brain Morphometry

Background:

• Quantitative evaluation of structural brain MRI is crucial for large-scale research and clinical applications.
• Existing pipelines can be time-consuming, necessitating faster and reliable alternatives.

Purpose of the Study:

• To evaluate deep learning models (DeepSCAN, FastSurferCNN, QuickNAT) for brain segmentation and cortex parcellation.
• To assess the performance of an 11-minute surface reconstruction pipeline using these models as input.
• To compare the pipeline's efficiency and reliability against established methods like FreeSurfer.

Main Methods:

• Three deep learning models (DeepSCAN, FastSurferCNN, QuickNAT) were used as input for a surface reconstruction pipeline.
• Performance was evaluated on large human MRI datasets and a synthetic dataset with known metrics.
• Evaluation criteria included agreement with FreeSurfer, reproducibility, age stability, contrast sensitivity, and synthetic metric accuracy.

Main Results:

• The DeepSCAN-based pipeline showed the highest agreement with FreeSurfer on human data.
• DeepSCAN demonstrated the greatest fidelity to expected metrics on the synthetic dataset.
• The pipeline achieved an 11-minute processing time, significantly faster than traditional methods.

Conclusions:

• The DeepSCAN-based surface reconstruction pipeline is a rapid and reliable alternative for structural MRI processing.
• Its efficiency and reliability make it suitable for high-throughput research applications.
• Further studies are needed to evaluate its robustness, pathological variability, and clinical diagnostic utility.