Abstract
White matter hyperintensities (WMH) are commonly found in the brains of healthy elderly individuals and have been associated with various neurological and geriatric disorders. In this paper, we present a study using deep fully convolutional network and ensemble models to automatically detect such WMH using fluid attenuation inversion recovery (FLAIR) and T1 magnetic resonance (MR) scans. The algorithm was evaluated and ranked 1st in the WMH Segmentation Challenge at MICCAI 2017. In the evaluation stage, the implementation of the algorithm was submitted to the challenge organizers, who then independently tested it on a hidden set of 110 cases from 5 scanners. Averaged dice score, precision and robust Hausdorff distance obtained on held-out test datasets were 80%, 84% and 6.30mm respectively. These were the highest achieved in the challenge, suggesting the proposed method is the state-of-the-art. In this paper, we provide detailed descriptions and quantitative analysis on key components of the system. Furthermore, a study of cross-scanner evaluation is presented to discuss how the combination of modalities and data augmentation affect the generalization capability of the system. The adaptability of the system to different scanners and protocols is also investigated. A quantitative study is further presented to test the effect of ensemble size. Additionally, software and models of our method are made publicly available. The effectiveness and generalization capability of the proposed system show its potential for real-world clinical practice.
Introduction
White Matter Hyperintensities (WMH) are brain lesions frequently observed in aging populations and are strongly linked to neurological conditions such as stroke, dementia, and cognitive decline. Quantifying these lesions is essential for tracking disease progression, yet manual segmentation by neuroradiologists is a labor-intensive process prone to significant inter-observer variability.
Automated segmentation faces several hurdles: lesions are often small, irregularly shaped, and scattered throughout the white matter. Moreover, MRI data varies significantly across different scanners and protocols, leading to differences in resolution, voxel size, and signal intensity. A robust automated system must generalize across these variations to be clinically useful.
Methodology
The core of the proposed system is a Fully Convolutional Network (FCN) based on the U-Net architecture. The authors made several strategic design choices to optimize performance for WMH segmentation:
- 2D Architecture: The model processes 2D axial slices rather than 3D volumes. This choice was driven by the high variability in slice thickness across datasets, where 2D slices provided more consistent texture information.
- Multi-Modal Input: The network utilizes two-channel inputs consisting of FLAIR and T1-weighted MRI scans. While FLAIR is the primary modality for WMH, T1 provides complementary structural context.
- Modified U-Net: The architecture features 19 convolutional layers. Standard 3x3 kernels in the initial layers were replaced with 5x5 kernels to capture broader spatial dependencies.
- Dice Loss Optimization: To address the extreme class imbalance (lesions occupy a tiny fraction of the brain), the model was trained using a Dice loss function, which directly optimizes for spatial overlap.
- Ensemble Learning: The final output is an average of probability maps from an ensemble of three identical U-Nets, each trained with different initializations and data shuffling to improve robustness and reduce variance.
Experimental Setup
The system was developed and validated using the MICCAI 2017 WMH Segmentation Challenge dataset. This dataset is particularly challenging due to its multi-center, multi-scanner nature.
The training set comprised 60 cases from three scanners (Utrecht, Singapore, and GE3T). The hidden test set included 110 cases from five scanners, including two (GE1.5T and PETMR) that were entirely unseen during the training phase. This setup rigorously tested the model's ability to generalize to new clinical environments.
Results and Discussion
The proposed ensemble method achieved the top rank in the MICCAI 2017 Challenge. On the hidden test set, it reached an average Dice Similarity Coefficient (DSC) of 80% and a lesion recall of 84%.
Key Insights:
- Cross-Scanner Robustness: The model maintained high performance on unseen scanners (recalls of 90% and 84%), demonstrating that the combination of multi-modal data and ensemble averaging effectively mitigates scanner-specific biases.
- Efficiency: Despite using an ensemble, the system is highly efficient, processing a full 3D scan in approximately 8 seconds on a standard GPU.
- The Power of Ensembles: Ensembles were found to be critical for smoothing boundary predictions and preventing catastrophic errors on out-of-distribution data, providing a more reliable clinical tool than single-model approaches.
Conclusion
This study presents a state-of-the-art solution for automated WMH segmentation. By leveraging a 2D U-Net ensemble and multi-modal MRI data, the system achieves high accuracy and exceptional generalization across diverse scanner types. Future work will focus on improving the detection of very small lesions and incorporating inter-slice context, perhaps through recurrent architectures, to further refine segmentation boundaries. The public availability of the code and models supports further research and potential clinical integration.
Summary updated