Automatic segmentation of white matter hyperintensities in routine clinical brain MRI by 2D VB-Net: A large-scale study

Wenhao Zhu, Hao Huang, Yaqi Zhou, Feng Shi, Hong Shen, Ran Chen, Rui Hua, Wei Wang, Shabei Xu, Xiang LuoFrontiers in Aging Neuroscience 2022 · DOI

TL;DR

This paper introduces a novel 2D VB-Net deep learning model for automatically segmenting white matter hyperintensities (WMH) in routine clinical brain MRI scans. By leveraging multi-modal MRI data and a refined labeling protocol, their method achieves high accuracy and outperforms existing approaches, offering a robust and lightweight solution for clinical applications.

What Problem Does This Paper Address?

White matter hyperintensities (WMH) are common findings in brain MRIs, particularly in older adults, and are associated with various neurological disorders like cognitive impairment and stroke. Quantifying WMH is crucial for understanding disease progression and treatment efficacy. However, manual segmentation is time-consuming, prone to significant inter-observer variability, and thus impractical for large-scale studies or routine clinical use. Existing automated methods often struggle with the specific characteristics of routine clinical MRI data, such as thick-slice protocols and the presence of coexisting lesions, limiting their clinical applicability.

Methods

The core of this work is the 2D VB-Net, a convolutional neural network adapted from a successful 3D architecture. Key methodological contributions include:

  • 2D VB-Net Architecture: This network utilizes a 2D encoder-decoder structure with residual connections and bottleneck layers. It's designed to efficiently process 2D thick-slice MRI data, which is more common in clinical settings than thin-slice data used in many research studies. The network takes T1-weighted, T2-weighted, and FLAIR images as input to better differentiate WMH from other lesions.
  • Multi-Modality Labeling Protocol: To improve the accuracy and consistency of ground truth data, the researchers developed a semi-automated labeling protocol. This involves using a histogram-based threshold on FLAIR images, followed by manual correction and cross-checking across T1-weighted and T2-weighted images. This approach significantly reduced inter-observer variability, achieving a Dice consistency of 0.878.
  • Weighted Dice Loss: Recognizing that WMH regions are often sparse, a weighted Dice loss function was implemented. This assigns higher importance to "definite" WMH regions (agreed upon by both observers) and "suspected" WMH regions (identified by only one observer), helping the model focus on accurate segmentation of these critical areas.
  • WMH Subclassification: The model can further subclassify WMH into four categories (juxtaventricular, periventricular, deep, and juxtacortical) based on their location relative to the ventricles. This provides more detailed information about WMH distribution, which can be clinically relevant.

Datasets

  • In-house Dataset: A large dataset of 1,045 subjects with WMH was collected from Tongji Hospital. These subjects exhibited WMH of varying degrees and were scanned on seven different MRI scanners (3T and 1.5T) using 2D axial protocols. This dataset was split into:
  • Multi-scanner dataset (869 subjects) for training and cross-validation.
  • Two independent datasets (IDS 1: 102 subjects, IDS 2: 74 subjects) for validation.
  • A silver standard dataset (20 subjects) with labels generated by five neuroradiologists and merged by majority voting, used for rigorous evaluation.
  • 2017 MICCAI WMH Segmentation Challenge Dataset: This public dataset, consisting of 60 subjects with thin-slice (3mm) FLAIR and T1-weighted images, was used to validate the model's performance on a different data acquisition protocol.

Key Insights and Ideas

  • The Importance of Routine Clinical Data: The study highlights that many existing WMH segmentation methods are optimized for research-grade, thin-slice data. Their work demonstrates the necessity and feasibility of developing robust models for the more common, thick-slice 2D MRI protocols used in everyday clinical practice.
  • Multi-Modality is Key for Differentiation: Using T1-weighted, T2-weighted, and FLAIR images together significantly improves the ability to distinguish WMH from other lesions that might appear hyperintense on a single sequence, such as lacunes or cortical infarcts.
  • Semi-Automated Labeling Bridges the Gap: The refined labeling protocol, combining automated initial segmentation with expert review, dramatically improved inter-observer consistency. This is crucial for creating reliable ground truth data for training deep learning models.
  • Lightweight and Deployable: The 2D VB-Net is noted for its relatively small model size (1.08M parameters), making it lightweight and potentially easier to deploy in clinical settings or cloud-based applications.

Key Results

  • High Segmentation Accuracy: The 2D VB-Net achieved superior performance compared to state-of-the-art methods (uResNet, 3D V-Net, VGGNet) on the multi-scanner dataset, with Dice scores of 0.878 for WMH and 0.789 on the MICCAI challenge dataset.
  • Improved Inter-Observer Consistency: The proposed labeling protocol increased Dice consistency from 0.750 to 0.943 for WMH and from 0.538 to 0.878 when comparing manual delineations.
  • Robust Generalization: The model performed well on independent datasets and even on the thin-slice MICCAI challenge dataset, demonstrating its ability to generalize across different imaging protocols and scanners.
  • Clinical Correlation: The algorithm showed excellent consistency with manual delineations when correlating WMH volumes with visual rating scores (Fazekas scale), outperforming other competing methods.
  • Accurate Co-Lesion Segmentation: The model effectively segmented not only WMH but also other coexisting intracranial lesions, which is a significant advantage for clinical utility.

Limitations and Future Work

  • Segmentation Accuracy for Other Lesions: While the model performs well on WMH, the accuracy for other coexisting intracranial lesions (like PVS) is slightly lower. This is attributed to the more challenging nature of labeling these lesions, especially in thick-slice images.
  • Ground Truth Ambiguity: Despite efforts, there's still room for improvement to perfectly match the ground truth generated by experienced observers, particularly for WMH segmentation.
  • Dynamic Observations: The authors suggest that future work should include dynamic observations in large-scale studies to further understand the role of WMH in disease progression and validate the clinical value of their method.
  • Subclassification Limitations: While the subclassification is useful, the authors acknowledge that uniform standards for WMH distribution division are still evolving.

Summary updated

Abstract

White matter hyperintensities (WMH) are imaging manifestations frequently observed in various neurological disorders, yet the clinical application of WMH quantification is limited. In this study, we designed a series of dedicated WMH labeling protocols and proposed a convolutional neural network named 2D VB-Net for the segmentation of WMH and other coexisting intracranial lesions based on a large dataset of 1,045 subjects across various demographics and multiple scanners using 2D thick-slice protocols that are more commonly applied in clinical practice. Using our labeling pipeline, the Dice consistency of the WMH regions manually depicted by two observers was 0.878, which formed a solid basis for the development and evaluation of the automatic segmentation system. The proposed algorithm outperformed other state-of-the-art methods (uResNet, 3D V-Net and Visual Geometry Group network) in the segmentation of WMH and other coexisting intracranial lesions and was well validated on datasets with thick-slice magnetic resonance (MR) images and the 2017 medical image computing and computer assisted intervention WMH Segmentation Challenge dataset (with thin-slice MR images), all showing excellent effectiveness. Furthermore, our method can subclassify WMH to display the WMH distributions and is very lightweight. Additionally, in terms of correlation to visual rating scores, our algorithm showed excellent consistency with the manual delineations and was overall better than those from other competing methods. In conclusion, we developed an automatic WMH quantification framework for multiple application scenarios, exhibiting a promising future in clinical practice.
BibTeX
@inproceedings{zhu2022automatic,
  author = {Wenhao Zhu and Hao Huang and Yaqi Zhou and Feng Shi and Hong Shen and Ran Chen and Rui Hua and Wei Wang and Shabei Xu and Xiang Luo},
  title = {{Automatic segmentation of white matter hyperintensities in routine clinical brain MRI by 2D VB-Net: A large-scale study}},
  booktitle = {Frontiers in Aging Neuroscience},
  year = {2022},
  doi = {10.3389/fnagi.2022.915009},
}
Figure