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

This is the authors' abstract. We don't add key points for this paper.

Objective: The aim of our study was to assess the reproducibility of deep learning-based automatic segmentation of brain structures in MRI scans across different scanner types and magnetic field strengths, particularly focusing on the comparison between 1.5 T and 3 T MRI scanners. Methods: Our analysis encompassed a comprehensive examination of MRI images, focusing on the consistency of volumetric segmentation. We utilized advanced deep learning techniques with human-in-the-loop as a part of the workflow for segmenting brain structures and compared results across subsequent scans using the same and different scanner types. Results: Our findings revealed high consistency in volumetric segmentation when comparing scans conducted on the same type of scanner (1.5 T to 1.5 T or 3 T to 3 T). The study revealed slightly better segmentation results for 1.5 T scanners compared to 3 T scanners when each was used independently. However, cross-comparisons between different scanner types (1.5 T vs. 3 T) demonstrated slightly less consistency, highlighting the influence of magnetic field strength on segmentation accuracy. Conclusion: This study emphasizes the necessity of using the same scanner type and protocol for reliable MRI studies, particularly for brain atrophy monitoring. The high repeatability of deep learning-based segmentation under these conditions confirms its efficacy for clinical and research applications.