TL;DR
Researchers have developed the first comprehensive digital multi-atlas of the neonatal rabbit brain, crucial for preclinical research. This atlas, built from high-resolution MRI scans and detailed manual segmentations, enables automated brain region identification, accelerating studies on brain development and injury in this important animal model.
What Problem Does This Paper Address?
Translational research bridging animal models and human conditions often faces challenges due to anatomical differences. While rodents are common models, their brain development and structure differ significantly from humans. Rabbits, with their more similar perinatal brain development and white matter maturation timing, are increasingly used. However, analyzing their brains with MRI requires accurate segmentation of anatomical structures. Manual segmentation is time-consuming and prone to variability, while existing automated methods rely on atlases, and a detailed neonatal rabbit brain atlas was previously unavailable. This paper addresses this gap by creating a high-resolution, multi-modal digital atlas for the neonatal rabbit brain.
Methods
The core of this work is the creation of a multi-atlas for neonatal rabbit brains. The process involved:
- Data Acquisition: High-resolution T1-weighted and diffusion-weighted MRI (DWI) scans were acquired from 12 neonatal rabbit subjects.
- Stereotaxic Orientation: A novel stereotaxic orientation was proposed, adapting the bicommissural plane to a 45° angle with the horizontal plane, deemed more suitable for the developing rabbit brain than traditional rat-based orientations.
- Manual Segmentation: A rigorous manual segmentation process was undertaken, involving iterative refinement, label fusion, and manual adjustments. This was performed on a T1-weighted modality, with DWI data used for cross-validation and refinement.
- Atlas Propagation and Fusion: To build the multi-atlas, a semi-automatic approach was used. Initial segmentations were propagated to new subjects using affine and non-rigid registration. Multiple label fusion techniques (Majority Voting, STAPLE, STEPS) were employed to combine segmentations from different atlases, aiming to minimize bias and improve accuracy.
- Automatic Segmentation Algorithm: The developed multi-atlas was then used to create an automatic segmentation algorithm. This algorithm leverages the learned correspondences from the multi-atlas to segment new neonatal rabbit brains.
- Probabilistic Atlas Creation: A probabilistic atlas was generated by averaging the segmentations of all subjects, providing a representation of the average brain and the likelihood of tissue types at each voxel.
Datasets
The study utilized a newly created dataset of 12 neonatal rabbit brains. These brains were acquired using high-resolution micro-MRI (T1-weighted and DWI). This dataset forms the basis of the introduced multi-atlas and probabilistic atlas. The authors also mention preliminary in vivo experiments on four neonatal rabbits, which were used to test the algorithm's adaptability to different acquisition protocols.
Key Insights and Ideas
- Rabbit Brain Anatomy is Distinct: The authors highlight that standard rat-based stereotaxic orientations are not directly applicable to rabbit brains, necessitating a new approach.
- Iterative Refinement is Key: The process of iteratively adding subjects to the multi-atlas and refining segmentations significantly improved accuracy and reduced manual effort over time.
- Multi-Modal Fusion Enhances Accuracy: Combining information from different MRI modalities (T1 and DWI) in label fusion methods generally led to better segmentation results.
- Open-Source Approach is Crucial: Making the atlas and associated software publicly available is vital for reproducibility and for enabling future research in this area.
Key Results
- A multi-modal multi-atlas of the neonatal rabbit brain was created, parcellating 89 distinct anatomical regions across 12 subjects.
- The manual segmentation time decreased significantly from 20-25 hours for the first subject to 2-6 hours for subsequent subjects due to the iterative refinement process.
- The automatic segmentation algorithm, when run on a standard computer, took approximately 2 hours and 10 minutes to process all 12 subjects, including label fusion.
- Validation studies, including inter-operator variability assessment and leave-one-out cross-validation, demonstrated the robustness and accuracy of the proposed method, with high Dice scores and low distances between manual and automatic segmentations for many regions.
- The algorithm showed promising adaptability to in vivo acquisitions with different protocols, although multi-modal fusion was limited by lower DWI resolution in those cases.
Limitations and Future Work
- Bias from Under-represented Subjects: The authors acknowledge that the distribution of term/pre-term and male/female subjects might introduce bias in propagation to under-represented groups.
- Manual Refinement Still Necessary: While automated, the process still requires manual adjustments, especially for regions with less distinct boundaries or those requiring histological criteria.
- In Vivo Protocol Differences: The current multi-modal approach is less effective with in vivo DWI data due to lower resolution compared to ex vivo scans.
- Future Work: The authors propose making the atlas and software an open-source, version-controlled repository to facilitate ongoing improvements and adaptations for future studies. They also suggest exploring more sophisticated numerical methods for parameter tuning and further validation on diverse datasets.
Summary updated