Auto-context Convolutional Neural Network for Geometry-Independent Brain Extraction in Magnetic Resonance Imaging

Seyed Sadegh Mohseni Salehi, Deniz Erdogmus, Ali GholipourarXiv 2017 · arXiv

Abstract

Brain extraction, or "skull stripping," is a foundational step in neuroimaging pipelines. Traditional methods often rely on atlas registration or geometric assumptions, which fail in challenging scenarios like fetal MRI where brain orientation is arbitrary. We present Auto-Net, a deep learning framework that combines multi-view 2D convolutional neural networks (CNNs) with an iterative "auto-context" refinement loop. By learning intrinsic brain features rather than relying on templates, Auto-Net achieves state-of-the-art performance on adult benchmarks (LPBA40, OASIS) and demonstrates superior robustness in fetal brain extraction, significantly outperforming traditional tools.

Introduction

Isolating brain tissue from non-brain structures in MRI scans is critical for subsequent volumetric analysis and diagnosis. Standard tools like the Brain Extraction Tool (BET) or ROBEX frequently struggle when image registration fails or when dealing with non-standard geometries. This is particularly problematic in fetal MRI, where the fetus can be in any orientation and is surrounded by complex maternal tissues. To address these limitations, we propose a geometry-independent, registration-free approach that leverages the power of deep learning and iterative refinement.

Methods

The Auto-Net framework employs two primary architectures: a 2.5D Voxelwise CNN and an Auto-U-net. The 2.5D approach utilizes three parallel pathways to process 2D patches from orthogonal planes (axial, coronal, sagittal) at multiple scales, capturing both local texture and global context without the memory overhead of 3D convolutions. The Auto-U-net variant provides a faster, fully convolutional alternative for whole-image processing.

The defining feature of both models is the Auto-context algorithm. This iterative process uses the probability map from a previous step as an additional input channel for the next, allowing the network to refine its predictions based on surrounding spatial context. This effectively corrects "holes" or false negatives by learning that a voxel's likelihood of being brain tissue is highly dependent on its neighbors.

Figure 1: Overview of the Auto-Net architectures, including the 2.5D multi-view pathways (a, b), the U-net structure (c), and the iterative Auto-context refinement loop (d).

Results

The models were evaluated across three datasets: LPBA40 (healthy adults), OASIS (Alzheimer’s patients), and a challenging Fetal MRI dataset.

Adult Benchmarks

On the LPBA40 dataset, Auto-Net achieved a state-of-the-art Dice overlap coefficient of 97.73%, surpassing previous deep learning methods like PCNN. The iterative auto-context refinement consistently improved performance across all metrics compared to single-pass versions of the networks.

Table 6: Performance comparison on LPBA40 and OASIS datasets. Auto-Net variants consistently outperform traditional tools and previous deep learning models.

Fetal MRI Performance

The most significant gains were observed in fetal MRI, where traditional tools typically fail due to arbitrary brain orientations. Auto-Net achieved a Dice score of 95.97%, a substantial improvement over BET and 3dSkullStrip, which averaged between 80-83%. The orientation-agnostic nature of the learned features allowed Auto-Net to maintain high accuracy regardless of fetal position.

Discussion and Conclusion

Auto-Net provides a robust, efficient solution for brain extraction that is independent of brain geometry and orientation. The 2.5D approach offers a practical "efficiency hack" for capturing 3D spatial information, while the Auto-U-net variant enables rapid processing in clinical settings. Future work will explore multi-modal integration to further enhance robustness across diverse MRI sequences.

Summary updated

BibTeX
@misc{salehi2017auto,
  author = {Seyed Sadegh Mohseni Salehi and Deniz Erdogmus and Ali Gholipour},
  title = {{Auto-context Convolutional Neural Network for Geometry-Independent Brain Extraction in Magnetic Resonance Imaging}},
  year = {2017},
  eprint = {1703.02083},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1703.02083},
}
Figure