White matter hyperintensity and stroke lesion segmentation and differentiation using convolutional neural networks

Ricardo Guerrero, Chen Qin, Ozan Oktay, Christopher Bowles, L. Chen, Richard Joules, Robin Wolz, Maria C. Valdés-Hernández, David Alexander Dickie, Joanna M. Wardlaw, Daniel Rueckert

TL;DR

Researchers have developed uResNet, a convolutional neural network that doesn't just find hyperintensities (bright spots) in brain MRIs, but successfully differentiates between white matter hyperintensities (WMH) and stroke lesions. By combining a U-shaped architecture with residual learning and a clever patch-sampling strategy, this method outperforms existing tools and provides clinical measurements that align closely with expert radiologists.

What Problem Does This Paper Address?

In aging brains and those with small vessel disease, MRI scans (specifically FLAIR sequences) often show "bright spots" known as White Matter Hyperintensities (WMH). Quantifying these is vital for tracking cognitive decline and testing new treatments.

However, there’s a major hurdle: stroke lesions also appear as bright spots on the same scans. Most automated tools are "pathology-blind"—they can detect that something is abnormal, but they can't tell if it’s a chronic WMH or a stroke-related injury. Manual differentiation by experts is the gold standard, but it is incredibly slow, expensive, and prone to human fatigue. This paper introduces a task-specific, fully automated method to solve this "identity crisis" in neuroimaging.

Methods

The authors introduce uResNet, a 2D fully convolutional neural network. Its design is a hybrid of two heavy hitters in deep learning:

  • The U-Net Architecture: It uses an "analysis path" to capture context and a symmetric "synthesis path" to enable precise localization.
  • Residual Learning: Unlike the original U-Net, uResNet incorporates residual elements (ResEle). Instead of just learning a direct mapping, the network learns the "residual" or the difference, which eases gradient flow and allows for deeper, more efficient training.
  • Summation over Concatenation: To reduce model complexity, the skip connections use element-wise summation rather than the standard concatenation found in U-Nets.

A critical part of their method is the sampling strategy. To avoid "location bias" (where the network expects the pathology to always be in the center of an image patch), they implemented a random center shift. They sampled patches containing lesions but randomly offset the center, forcing the network to learn features rather than just spatial coordinates.

Datasets

The study utilized a dataset of 167 subjects from a hospital stroke service.

  • Diversity: Data was acquired using three different MRI protocols on a 1.5T scanner.
  • Ground Truth: All images were meticulously delineated by experts. WMH were segmented using a semi-automated multispectral technique (MCMxxxVI), and stroke lesions were manually/semi-automatically extracted.
  • Clinical Data: A subset of 126 subjects included rich clinical metadata, such as Fazekas scores (a visual rating of WMH severity) and risk factors like blood pressure and smoking status.

Key Insights and Ideas

  • The Sampling "Aha" Moment: The researchers found that simply feeding the network patches centered on lesions created a massive bias. By introducing a random spatial offset during training, they boosted the Dice score (a measure of overlap) for WMH and strokes significantly (from ~67 to ~70 and ~29 to ~40, respectively).
  • FLAIR is King: While the model can handle multi-modal inputs (like T1-weighted images or probability maps), the authors found that FLAIR images alone provided nearly identical performance. This is a win for clinical utility, as it reduces the need for complex, multi-sequence acquisitions.
  • The Class Imbalance Struggle: Stroke lesions are much rarer than WMH (there were ~14x more WMH voxels in the dataset). This makes stroke segmentation much more "oscillatory" and difficult for the network, highlighting a persistent challenge in medical AI: the "needle in a haystack" problem.
  • 2D vs. 3D: Despite the trend toward 3D CNNs, the authors argue for 2D convolutions. Clinical FLAIR scans often have large slice thicknesses, making 3D convolutions less intuitive and potentially less accurate than processing the brain slice-by-slice.

Key Results

  • Superior Accuracy: uResNet achieved a Mean Dice score of 69.5 for WMH, significantly outperforming established methods like DeepMedic (66.6), LPA (64.7), and LGA (41.0).
  • Volume Correlation: The correlation between uResNet’s automated volumes and expert volumes was remarkably high ($R^2 = 0.951$).
  • Clinical Validity: The automated WMH volumes correlated better with the Fazekas visual rating scale than even the expert-annotated volumes in some cases, suggesting the model is capturing the clinical "burden" of the disease very effectively.
  • Risk Factor Association: The model successfully replicated known clinical associations, such as the link between WMH volume and enlarged peri-vascular spaces (BGPVS).

Limitations and Future Work

While uResNet excels at WMH, the stroke segmentation remains a tougher nut to crack, with a Mean Dice score of 40.0. The authors attribute this to the high heterogeneity of strokes and the significant class imbalance in the training data.

Future directions include:

  • Integrating pre-activated residual units to further improve generalization.
  • Exploring more sophisticated ways to handle class imbalance beyond simple patch sampling.
  • Testing the architecture on larger, multi-center datasets to ensure the model generalizes across different scanner manufacturers and hospital settings.

Reference: Guerrero, R., et al. (2018). "White matter hyperintensity and stroke lesion segmentation and differentiation using convolutional neural networks." NeuroImage: Clinical.

Summary updated

Abstract

White matter hyperintensities (WMH) are a feature of sporadic small vessel disease also frequently observed in magnetic resonance images (MRI) of healthy elderly subjects. The accurate assessment of WMH burden is of crucial importance for epidemiological studies to determine association between WMHs, cognitive and clinical data; their causes, and the effects of new treatments in randomized trials. The manual delineation of WMHs is a very tedious, costly and time consuming process, that needs to be carried out by an expert annotator (e.g. a trained image analyst or radiologist). The problem of WMH delineation is further complicated by the fact that other pathological features (i.e. stroke lesions) often also appear as hyperintense regions. Recently, several automated methods aiming to tackle the challenges of WMH segmentation have been proposed. Most of these methods have been specifically developed to segment WMH in MRI but cannot differentiate between WMHs and strokes. Other methods, capable of distinguishing between different pathologies in brain MRI, are not designed with simultaneous WMH and stroke segmentation in mind. Therefore, a task specific, reliable, fully automated method that can segment and differentiate between these two pathological manifestations on MRI has not yet been fully identified. In this work we propose to use a convolutional neural network (CNN) that is able to segment hyperintensities and differentiate between WMHs and stroke lesions. Specifically, we aim to distinguish between WMH pathologies from those caused by stroke lesions due to either cortical, large or small subcortical infarcts. The proposed fully convolutional CNN architecture, called uResNet, that comprised an analysis path, that gradually learns low and high level features, followed by a synthesis path, that gradually combines and up-samples the low and high level features into a class likelihood semantic segmentation. Quantitatively, the proposed CNN architecture is shown to outperform other well established and state-of-the-art algorithms in terms of overlap with manual expert annotations. Clinically, the extracted WMH volumes were found to correlate better with the Fazekas visual rating score than competing methods or the expert-annotated volumes. Additionally, a comparison of the associations found between clinical risk-factors and the WMH volumes generated by the proposed method, was found to be in line with the associations found with the expert-annotated volumes.
BibTeX
@inproceedings{guerrero2017white,
  author = {Ricardo Guerrero and Chen Qin and Ozan Oktay and Christopher Bowles and L. Chen and Richard Joules and Robin Wolz and Maria C. Valdés-Hernández and David Alexander Dickie and Joanna M. Wardlaw and Daniel Rueckert},
  title = {{White matter hyperintensity and stroke lesion segmentation and differentiation using convolutional neural networks}},
  booktitle = {NeuroImage},
  year = {2017},
  doi = {10.1016/j.nicl.2017.12.022},
}

Loading related content...

Figure