Related Experiment Video
Updated: Nov 4, 2025
04:25
Author Spotlight: Bridging Gaps in Anatomy and Establishing a Foundation for Algorithmic Studies
Published on: December 15, 2023
3.2K
Generalizing Deep Whole Brain Segmentation for Pediatric and Post- Contrast MRI with Augmented Transfer Learning.
Camilo Bermudez1, Justin Blaber2, Samuel W Remedios3
1Department of Biomedical Engineering, Vanderbilt University, 2201 West End Ave, Nashville, TN, USA 37235.
Summary
Data augmentation enhances deep learning models like SLANT for brain MRI segmentation, improving accuracy across diverse datasets. This method overcomes limitations of traditional transfer learning, preventing performance degradation on original data.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuroscience
Background:
- Deep neural networks face generalizability challenges in clinical magnetic resonance imaging (MRI) due to data variability.
- The Spatially Localized Atlas Network Tiles (SLANT) effectively segments brain MRI but requires enhanced generalizability for multi-site studies.
- Transfer learning (TL) can adapt models but risks performance degradation on original datasets.
Purpose of the Study:
- To explore data augmentation as a method to improve the generalizability of the SLANT approach for brain MRI segmentation.
- To adapt SLANT for anatomical variations (pediatric vs. adult) and scanning protocols (contrast-enhanced vs. non-contrast T1w MRI).
- To compare data augmentation with traditional transfer learning in terms of performance and generalizability.
Main Methods:
- Two datasets were used: pediatric T1w MRI (n=30) and paired pre-/post-contrast clinical T1w MRI (n=36).
- SLANT's transfer learning step was augmented with either new data only or both original and new data.
- Segmentation accuracy was assessed using Dice Similarity Coefficient (DSC) against manually corrected labels or prior automated segmentations.
Main Results:
- Both data augmentation approaches significantly improved performance over baseline SLANT for pediatric (0.89 vs. 0.82 DSC) and contrast-enhanced MRI (0.80 vs. 0.76 DSC).
- Strict transfer learning (new data only) led to decreased performance on the original test set.
- Data augmentation proved superior to strict transfer learning, preserving performance on original data while improving adaptation.
Conclusions:
- Data augmentation is a superior strategy to strict transfer learning for enhancing the generalizability of SLANT for brain MRI segmentation.
- This approach effectively adapts SLANT to anatomical and protocol variations without compromising performance on existing datasets.
- The findings support broader application of volumetric MRI assessment in multi-site and diverse clinical studies.
