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

Abstract

Accurate assessment of white matter hyperintensity (WMH) burden is essential for epidemiological studies investigating the links between small vessel disease, cognitive decline, and clinical outcomes. However, manual delineation is labor-intensive and complicated by the presence of stroke lesions, which also appear hyperintense on MRI. This paper presents a novel convolutional neural network (CNN) architecture, uResNet, designed to simultaneously segment and differentiate WMHs from stroke lesions (including cortical, large subcortical, and small subcortical infarcts). The architecture utilizes an analysis-synthesis path with residual learning to capture multi-scale features. Experimental results demonstrate that uResNet outperforms state-of-the-art methods in segmentation accuracy and produces volumes that correlate strongly with expert annotations and clinical Fazekas scores.

Introduction

In neuroimaging, particularly within aging populations, Fluid-Attenuated Inversion Recovery (FLAIR) sequences frequently reveal "bright spots" known as White Matter Hyperintensities (WMH). These regions are critical biomarkers for small vessel disease and cognitive impairment. A significant challenge in automated quantification is that stroke lesions exhibit similar hyperintense signals, leading to misclassification by "pathology-blind" algorithms.

Distinguishing between these two types of pathologies is vital for clinical accuracy. While manual segmentation by expert radiologists remains the gold standard, it is time-consuming and subject to inter-observer variability. This research introduces a fully automated method to solve this differentiation task, marking one of the first explicit attempts to separate WMH from various stroke-related pathologies using deep learning.

Methodology

The uResNet Architecture

The proposed model, uResNet, is a 2D fully convolutional neural network that builds upon the classic U-Net framework. It consists of two primary components:

  1. Analysis Path: This path gradually extracts low- and high-level features through a series of convolutional layers.
  2. Synthesis Path: This path combines and up-samples these features to generate a class-likelihood semantic segmentation map.

To enhance performance, the authors integrated residual elements (ResEle). Unlike standard U-Nets that learn direct mappings, uResNet learns the residual difference, which facilitates gradient flow and allows for more efficient training of deeper networks. Additionally, the model uses element-wise summation in its skip connections rather than concatenation, reducing the overall number of parameters and computational complexity.

Spatial Sampling Strategy

A key innovation in this work is the addressing of "location bias." In many medical imaging tasks, pathologies tend to be centered in training patches, leading the network to rely on spatial coordinates rather than morphological features. To counter this, the researchers implemented a random center shift during patch sampling. By randomly offsetting the center of the lesion within the training patch, the network was forced to learn the underlying characteristics of the hyperintensities, significantly improving the Dice similarity coefficient (DSC) for both WMH and stroke lesions.

Experimental Setup

The model was trained and validated using a dataset of 167 subjects from a hospital stroke service. The data represented a real-world clinical environment, acquired using three different MRI protocols on a 1.5T scanner.

Ground truth was established through meticulous expert delineation. WMHs were segmented using a semi-automated multispectral technique, while stroke lesions were manually extracted. Clinical validation was performed on a subset of 126 subjects, comparing automated volumes against Fazekas visual rating scores and known clinical risk factors such as blood pressure and smoking status.

Results and Discussion

Segmentation Performance

uResNet achieved a Mean Dice score of 69.5 for WMH, outperforming established algorithms including DeepMedic (66.6), LPA (64.7), and LGA (41.0). The correlation between the automated volumes and expert-derived volumes was exceptionally high, with an $R^2$ value of 0.951.

Clinical Correlation

The clinical utility of the model was confirmed through several metrics:

  • Fazekas Scale: The automated WMH volumes showed a stronger correlation with Fazekas visual ratings than the expert-annotated volumes in certain instances, suggesting the model effectively captures the clinical burden of the disease.
  • Risk Factors: The model successfully replicated established clinical associations, such as the relationship between WMH volume and enlarged peri-vascular spaces (BGPVS).
  • Modality Efficiency: Interestingly, the researchers found that using FLAIR images alone yielded results nearly identical to multi-modal inputs (T1, etc.). This suggests that the model can be deployed in clinical settings where only standard FLAIR sequences are available.

Challenges in Stroke Differentiation

While the model excelled at WMH segmentation, stroke lesion segmentation proved more difficult, yielding a Mean Dice score of 40.0. This is attributed to the extreme class imbalance (WMH voxels outnumbered stroke voxels by 14:1) and the high morphological heterogeneity of stroke injuries.

Conclusion

uResNet represents a significant step forward in automated neuroimaging analysis. By successfully differentiating between WMH and stroke lesions, the model provides a more nuanced and clinically relevant assessment of brain health than previous "pathology-blind" tools. Future work will focus on addressing class imbalance through advanced sampling techniques and exploring pre-activated residual units to further refine segmentation accuracy for rare or heterogeneous pathologies.

Summary updated

Abstract

The accurate assessment of White matter hyperintensities (WMH) burden is of crucial importance for epidemiological studies to determine association between WMHs, cognitive and clinical data. The manual delineation of WMHs is tedious, costly and time consuming. This is further complicated by the fact that other pathological features (i.e. stroke lesions) often also appear as hyperintense. Several automated methods aiming to tackle the challenges of WMH segmentation have been proposed, however cannot differentiate between WMH and strokes. Other methods, capable of distinguishing between different pathologies in brain MRI, are not designed with simultaneous WMH and stroke segmentation in mind. 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. As far as we know, this is the first time such differentiation task has explicitly been proposed. The proposed fully convolutional CNN architecture, is comprised of 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. Additionally, a comparison of the associations found between clinical risk-factors and the WMH volumes generated by the proposed method, were 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},
}
Figure