Enhancing Hierarchical Transformers for Whole Brain Segmentation with Intracranial Measurements Integration
Xin Yu1, Yucheng Tang2,3, Qi Yang1
1Computer Science, Vanderbilt University, Nashville, TN, USA.
Proceedings of Spie--The International Society for Optical Engineering
|September 2, 2024
Summary
This study enhances whole brain segmentation using deep learning, accurately measuring total intracranial volume (TICV) and posterior fossa volume (PFV) with limited data. The UNesT model achieves precise estimation for these key metrics alongside 132 brain regions.
Area of Science:
- Neuroimaging
- Medical Image Analysis
- Artificial Intelligence
Background:
- Whole brain segmentation via MRI is crucial for non-invasive brain region measurement.
- Existing methods lack comprehensiveness for intracranial measurements like total intracranial volume (TICV) and posterior fossa volume (PFV).
- Deep learning for intracranial measurements is hindered by limited annotated data.
Purpose of the Study:
- To enhance whole brain segmentation using the UNesT hierarchical transformer.
- To simultaneously segment 133 whole brain classes, including TICV and PFV.
- To address data scarcity challenges in deep learning for neuroimaging.
Main Methods:
- Pretraining the UNesT model on 4859 T1-weighted 3D volumes from diverse sites.
- Utilizing a multi-atlas segmentation pipeline for initial label generation.
- Fine-tuning the model on 45 OASIS dataset volumes with complete whole brain and TICV/PFV labels.
Main Results:
- The enhanced UNesT model achieves precise estimation of TICV and PFV.
- Performance for 132 brain regions remains comparable to existing methods.
- Evaluation using Dice similarity coefficients (DSC) confirms accuracy.
Conclusions:
- The proposed method effectively integrates TICV and PFV segmentation into whole brain analysis.
- This approach overcomes data limitations for deep learning in intracranial measurements.
- The UNesT model offers a robust solution for comprehensive brain structure analysis.
