Accurate brain extraction using Active Shape Model and Convolutional Neural Networks

Nguyen Ho Minh Duy, Manh-Duy Nguyen, Mai Thanh Nhat Truong, Pham The Bao, Thanh Binh NguyenarXiv 2018 · arXiv · OpenReview

TL;DR

This paper introduces ASMCNN, a novel brain extraction (skull stripping) method for MRI scans that combines Active Shape Models (ASM) and Convolutional Neural Networks (CNNs). By processing 2D sagittal slices and intelligently grouping them based on shape, ASMCNN achieves state-of-the-art accuracy, particularly for challenging cases with small or complex brain regions.

What Problem Does This Paper Address?

Accurately segmenting the brain from surrounding tissues (skull stripping) is a crucial preprocessing step in neuroimaging analysis. However, this task is notoriously difficult due to variations in image acquisition, the presence of abnormalities, and the inherent complexity and variability of brain anatomy. Existing methods often struggle with small or unusually shaped brain regions, leading to inaccurate segmentations that can compromise downstream analyses.

Methods

ASMCNN tackles brain extraction by leveraging a hybrid approach:

  • 2D Sagittal Plane Processing: Instead of working with full 3D volumes, ASMCNN processes 2D slices in the sagittal plane. This allows for more efficient processing and the development of specialized strategies.
  • Intelligent Grouping: Images are divided into three groups (Group I, II, and III) based on the shape and appearance of brain boundaries. This "divide-and-conquer" strategy allows for tailored processing for each group.
  • Active Shape Models (ASM): For groups with more regular brain shapes (Group II and III), a modified ASM is used to generate initial brain boundary estimates. ASM is effective at capturing general shape information.
  • Convolutional Neural Networks (CNNs): For Group I, which contains small and complex brain regions, a CNN is directly applied to segment pixels. For Groups II and III, CNNs are used to refine the ASM-generated boundaries, incorporating both local features and global spatial information.
  • Post-processing: Conditional Random Fields (CRFs), Gaussian processes, and custom rules are employed to further refine segmentation contours and correct errors, ensuring smooth and accurate final masks.

Datasets

The authors evaluated their method on three publicly available datasets:

  • IBSR (Internet Brain Segmentation Repository)
  • LPBA (LONI Probabilistic Brain Atlas)
  • OASIS (Open Access Series of Imaging Studies)

Key Insights and Ideas

  • The power of 2D processing for 3D problems: By cleverly processing 2D sagittal slices and then combining the results, the authors demonstrate that a 2D-centric approach can achieve excellent results, potentially overcoming limitations of 3D methods when data is scarce.
  • Grouping is key for robustness: Dividing images into groups based on shape allows for specialized algorithms, significantly improving performance on challenging cases (like small brains in Group I) where a single model would fail.
  • ASM + CNN synergy: Combining the global shape prior of ASM with the local feature extraction power of CNNs creates a robust pipeline that benefits from both approaches.
  • Exploiting symmetry: The sagittal plane's inherent symmetry is leveraged to improve prediction accuracy and develop more effective segmentation rules.

Key Results

ASMCNN demonstrated state-of-the-art performance across the evaluated datasets (IBSR, LPBA, and OASIS).

  • Quantitative: The paper reports superior Dice overlap and Jaccard index scores compared to seven other methods, with particularly high accuracy on the OASIS dataset. For example, on the OASIS dataset, ASMCNN achieved a Dice score of 95.07% ± 0.054, outperforming other methods.
  • Qualitative: ASMCNN produced smooth boundaries and accurately included gray and dura matter, outperforming methods like ROBEX and BEAST, which sometimes over- or under-segmented complex regions. It showed particular strength in handling small brain regions where other methods often failed.

Limitations and Future Work

  • Annotation Effort: ASMCNN requires more annotation effort than some end-to-end deep learning methods, needing additional data for group classification and bounding box extraction.
  • Hyperparameter Tuning: Users need to adjust certain hyperparameters for optimal performance.
  • Sequential Processing: The method's sequential nature (ASM before CNN) makes it slower than end-to-end deep learning approaches.
  • Future Work: The authors suggest investigating deep learning techniques for better shape and appearance models, exploring whole 3D volume processing, and generalizing their CNN architecture for 3D inputs. They also propose enhancing feature learning and exploring transfer learning.

Summary updated

Abstract

Brain extraction (skull stripping) is a challenging problem in neuroimaging. It is due to the variability in conditions from data acquisition or abnormalities in images, making brain morphology and intensity characteristics changeable and complicated. In this paper, we propose an algorithm for skull stripping in Magnetic Resonance Imaging (MRI) scans, namely ASMCNN, by combining the Active Shape Model (ASM) and Convolutional Neural Network (CNN) for taking full of their advantages to achieve remarkable results. Instead of working with 3D structures, we process 2D image sequences in the sagittal plane. First, we divide images into different groups such that, in each group, shapes and structures of brain boundaries have similar appearances. Second, a modified version of ASM is used to detect brain boundaries by utilizing prior knowledge of each group. Finally, CNN and post-processing methods, including Conditional Random Field (CRF), Gaussian processes, and several special rules are applied to refine the segmentation contours. Experimental results show that our proposed method outperforms current state-of-the-art algorithms by a significant margin in all experiments.
BibTeX
@misc{duy2018accurate,
  author = {Nguyen Ho Minh Duy and Manh-Duy Nguyen and Mai Thanh Nhat Truong and Pham The Bao and Thanh Binh Nguyen},
  title = {{Accurate brain extraction using Active Shape Model and Convolutional Neural Networks}},
  year = {2018},
  eprint = {1802.01268},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1802.01268},
}
Figure