Automatic Skull Stripping of Rat and Mouse Brain MRI Data Using U-Net

Li-Ming Hsu, Shuai Wang, Paridhi Ranadive, Woomi Ban, Tzu-Hao Harry Chao, Sheng Song, Domenic H. Cerri, Lindsay R. Walton, Margaret Broadwater, Sung-Ho Lee, Dinggang Shen, Yen-Yu I. ShihFrontiers in Neuroscience 2020 · DOI

Abstract

Accurate removal of magnetic resonance imaging (MRI) signal outside the brain, known as skull stripping, is a critical step in neuroimaging pipelines. In rodent studies, this is often performed manually, which is time-consuming and prone to operator bias. This study proposes a deep-learning framework based on the U-Net architecture to automate rodent brain segmentation. The method was validated using both in-house and public datasets, demonstrating superior performance over traditional rule-based methods across various MRI protocols, including T2-weighted and T2*-weighted imaging.

Introduction

In preclinical neuroscience, isolating brain tissue from non-brain signals—such as muscle, skin, and bone—is a prerequisite for downstream analysis. While automated tools for human MRI are well-established, rodent MRI presents unique challenges: smaller brain-to-scalp distances, distinct tissue geometries, and often lower image resolution. Consequently, many laboratories still rely on manual masking, a process that is labor-intensive and difficult to replicate across studies. This paper introduces a reliable, automated solution using the U-Net convolutional neural network to replace this manual bottleneck.

Methodology: The U-Net Framework

The authors employed U-Net, a convolutional neural network (CNN) architecture specifically designed for biomedical image segmentation. The network features a symmetric "U" shape: an encoder path that captures contextual features and a decoder path that enables precise localization of brain boundaries.

The U-Net architecture used for rodent brain segmentation, featuring an encoder-decoder structure with skip connections to preserve spatial information.

Technical highlights of the implementation include:

  • 2D Slice-by-Slice Processing: The model processes 2D axial slices, making it computationally efficient enough to run on standard hardware without requiring massive 3D computing resources.
  • Dice Coefficient Loss: A Dice-based loss function was used to handle the class imbalance inherent in brain imaging, where the target brain tissue occupies only a fraction of the total image volume.
  • Normalization: Rigorous spatial and intensity normalization were implemented to ensure the model's robustness across different scanners and protocols.

Datasets and Training

To ensure generalizability, the model was trained and validated on a diverse set of data:

  • CAMRI Dataset (In-house): Scans from 132 rats and 16 mice, covering high-resolution anatomical (T2w RARE) and functional (T2*w EPI) protocols.
  • OpenNeuro Dataset (Online): Publicly available data from 24 rats and 36 mice used to test the model on "unseen" data from different institutions and hardware.

Performance and Robustness

The U-Net approach consistently outperformed existing state-of-the-art tools, including RATS, PCNN, and SHERM. In rats, the model achieved a Dice similarity coefficient of 0.97 for anatomical scans and 0.96 for functional scans.

Comparison of Positive Predictive Value (PPV) across different methods, highlighting U-Net's high precision and stability relative to traditional tools like PCNN and SHERM.

Key advantages identified include:

  • Stability on EPI Data: While traditional methods often struggle with the artifacts and lower resolution of functional MRI (EPI), U-Net remained remarkably stable.
  • Handling Edge Cases: U-Net successfully identified challenging regions like the brainstem and olfactory bulbs, where rule-based methods frequently fail due to weak signal contrast.
  • Parameter-Free Operation: Unlike traditional tools that require manual parameter tuning (such as expected brain volume) for each study, the trained U-Net model adapts to new data automatically.

Discussion and Future Work

While highly effective for T2-weighted images, the current model's performance on other contrasts like T1-weighted images may require additional training data. Future developments could include 3D U-Net architectures to leverage inter-slice information for even smoother boundaries and multi-class segmentation to automatically delineate specific brain regions or nuclei.

Conclusion

The proposed U-Net framework provides a robust, automated, and bias-free tool for rodent brain MRI preprocessing. By eliminating the manual workload of skull stripping, this approach streamlines preclinical neuroimaging pipelines and enhances the reproducibility of research findings across the neuroscience community.

Source: Hsu et al. (2020). Automatic Skull Stripping of Rat and Mouse Brain MRI Data Using U-Net. Frontiers in Neuroscience. Link to Paper

Summary updated

Abstract

Accurate removal of magnetic resonance imaging (MRI) signal outside the brain, a.k.a., skull stripping, is a key step in the brain image pre-processing pipelines. In rodents, this is mostly achieved by manually editing a brain mask, which is time-consuming and operator dependent. Automating this step is particularly challenging in rodents as compared to humans, because of differences in brain/scalp tissue geometry, image resolution with respect to brain-scalp distance, and tissue contrast around the skull. In this study, we proposed a deep-learning-based framework, U-Net, to automatically identify the rodent brain boundaries in MR images. The U-Net method is robust against inter-subject variability and eliminates operator dependence. To benchmark the efficiency of this method, we trained and validated our model using both in-house collected and publicly available datasets. In comparison to current state-of-the-art methods, our approach achieved superior averaged Dice similarity coefficient to ground truth T2-weighted rapid acquisition with relaxation enhancement and T2<sup>∗</sup>-weighted echo planar imaging data in both rats and mice (all <i>p</i> < 0.05), demonstrating robust performance of our approach across various MRI protocols.
BibTeX
@inproceedings{hsu2020automatic,
  author = {Li-Ming Hsu and Shuai Wang and Paridhi Ranadive and Woomi Ban and Tzu-Hao Harry Chao and Sheng Song and Domenic H. Cerri and Lindsay R. Walton and Margaret Broadwater and Sung-Ho Lee and Dinggang Shen and Yen-Yu I. Shih},
  title = {{Automatic Skull Stripping of Rat and Mouse Brain MRI Data Using U-Net}},
  booktitle = {Frontiers in Neuroscience},
  year = {2020},
  doi = {10.3389/fnins.2020.568614},
}
Figure