Bayesian convolutional neural network based MRI brain extraction on nonhuman primates

Fang Liu, Gengyan Zhao, Fang Liu 0005, Jonathan A. Oler, Mary E. Meyerand, Ned H. Kalin, Rasmus M. BirnNeuroImage 2018 · DOI

TL;DR

This paper introduces a novel deep learning framework, Bayesian SegNet combined with a 3D Conditional Random Field (CRF), for highly accurate brain extraction in nonhuman primate MRI scans. This method significantly outperforms existing tools, offering a crucial advancement for nonhuman primate neuroimaging research by automating a previously labor-intensive process.

What Problem Does This Paper Address?

Brain extraction, or "skull stripping," is a fundamental preprocessing step in MRI analysis, essential for accurate registration and subsequent analyses like cortical thickness estimation. While current automated methods perform well on human brains, they struggle with nonhuman primates due to their smaller brain size, complex tissue structures, and larger, more prominent eye orbits filled with adipose tissue. This difficulty often necessitates manual intervention, which is time-consuming and prone to variability. The paper aims to develop a robust, automated solution for nonhuman primate brain extraction that matches or exceeds human brain extraction accuracy.

Methods

The core of the proposed method is Bayesian SegNet, a deep convolutional neural network (CNN) that incorporates Monte Carlo dropout during both training and testing. This Bayesian approach allows the network not only to segment the brain but also to quantify its uncertainty for each prediction. This uncertainty estimation is crucial for identifying potentially unreliable segmentations.

Following the Bayesian SegNet segmentation, a fully connected 3D Conditional Random Field (CRF) is employed as a post-processing step. This CRF leverages the 3D anatomical context of the entire brain volume to refine the probabilistic outputs from SegNet, considering relationships between voxels based on their location and intensity. This combination aims to improve segmentation accuracy by incorporating global spatial information.

Datasets

The study utilized a dataset of 100 periadolescent rhesus macaque MRI scans. These scans were collected using a 3T MRI scanner and were manually extracted by trained image scientists using the software SPAMALIZE. This dataset serves as the ground truth for training and evaluating the proposed method and its competitors.

Key Insights and Ideas

  • Uncertainty as a Confidence Measure: The Bayesian nature of SegNet, particularly through Monte Carlo dropout, provides a valuable measure of prediction uncertainty. This is critical for identifying areas where the model is less confident, potentially indicating challenging anatomical regions or data inconsistencies.
  • 3D Context is Crucial for Refinement: While SegNet provides 2D segmentation, the subsequent 3D CRF step significantly improves accuracy by considering the entire brain volume, effectively smoothing out local errors and enforcing global consistency.
  • Nonhuman Primate Challenges are Unique: The paper highlights specific anatomical features of nonhuman primate brains (e.g., large orbits with adipose tissue, sharp frontal lobe curvature) that pose significant challenges for existing brain extraction algorithms.
  • Deep Learning's Adaptability: The study demonstrates that deep learning methods, when properly designed and trained, can overcome these specific challenges, achieving superior performance compared to traditional methods.

Key Results

The proposed method, BSegNetCRF, achieved state-of-the-art performance on the nonhuman primate dataset.

  • It outperformed six popular publicly available brain extraction packages and three well-established deep learning methods.
  • The mean Dice coefficient was 0.985, and the mean average symmetric surface distance was 0.220 mm.
  • Statistical tests confirmed the superiority of BSegNetCRF over all compared methods (p < 10⁻⁴, Bonferroni corrected).
  • The maximum voxel-labeling uncertainty for Bayesian SegNet had a mean value of 0.116 across all subjects, with higher uncertainty observed at brain edges.
  • The processing time for a single subject was approximately 2 minutes, making it comparable to other edge-detecting methods and significantly faster than template-based methods.

Limitations and Future Work

  • Parameter Selection: The study did not perform subject-specific parameter selection or a grid search for parameters, which could potentially further improve performance but would be computationally prohibitive.
  • Limited Age Spectrum: The study focused on periadolescent rhesus monkeys. Future work will investigate the method's performance across a wider age spectrum and explore transfer learning for different age groups.
  • Further Data Modalities: The authors suggest future work could incorporate third-dimensional information and image noise information to further enhance brain extraction accuracy.
  • Uncertainty Behavior: While the study explored the influence of training set size and label inconsistency on uncertainty, a more thorough investigation into the stochastic nature of training and its impact on uncertainty is suggested.

Summary updated

BibTeX
@article{liu2018bayesian,
  author = {Fang Liu and Gengyan Zhao and Fang Liu and Jonathan A. Oler and Mary E. Meyerand and Ned H. Kalin and Rasmus M. Birn},
  title = {{Bayesian convolutional neural network based MRI brain extraction on nonhuman primates}},
  journal = {NeuroImage},
  year = {2018},
  doi = {10.1016/j.neuroimage.2018.03.065},
}
Figure