Generalizing deep whole-brain segmentation for post-contrast MRI with transfer learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroimaging

Background:

• Deep neural networks face generalizability challenges in clinical MRI due to data variability.
• The spatially localized atlas network tiles (SLANT) effectively segments non-contrast T1w MRI but requires adaptation for clinical data.
• Transfer learning (TL) can adapt models but may degrade performance on original datasets.

Purpose of the Study:

• To explore TL using unlabeled clinical data to improve SLANT's generalizability to scanning protocol variations.
• To adapt whole-brain segmentation for heterogeneous clinical T1w MRI data.
• To address performance degradation risks associated with TL in deep learning models.

Main Methods:

• Leveraged 480 unlabeled T1w MRI pairs (pre- and post-contrast) for TL.
• Used labels from pre-contrast images to train on post-contrast images within a five-fold cross-validation framework.
• Validated the adapted SLANT model on a separate test set of 29 paired scans.

Main Results:

• Improved reproducibility across imaging pairs, measured by reproducibility Dice coefficient (rDSC), from 0.72 to 0.82 compared to original SLANT.
• Significantly outperformed FreeSurfer v6.0.1 segmentation pipeline.
• Reduced root-mean-squared error of hippocampal volumetric estimates by 67% between paired images.

Conclusions:

• Demonstrated a pipeline using unlabeled clinical data to translate research-optimized algorithms for clinical use.
• Successfully adapted SLANT to generalize to heterogeneous clinical MRI acquisitions.
• Showcased the effectiveness of TL in enhancing deep learning model robustness for medical imaging applications.