Segmentation of Cerebral Small Vessel Diseases-White Matter Hyperintensities Based on a Deep Learning System

Wei Shan, Yunyun Duan, Yu Zheng, Zhenhua Wu, Wei Sheng Chan, Qun Wang, Peiyi Gao, Yaou Liu, Kunlun He, Yongjun WangFrontiers in Medicine 2021 · DOI

TL;DR

This paper introduces a novel deep learning system for automatically segmenting White Matter Hyperintensities (WMHs) from FLAIR MRI scans. The system achieves high accuracy and clinical acceptance, offering a faster and more objective alternative to manual segmentation for assessing Cerebral Small Vessel Diseases (CSVD).

What Problem Does This Paper Address?

Cerebral Small Vessel Diseases (CSVD) are a significant cause of neurological disorders, and White Matter Hyperintensities (WMHs) are a key indicator. Accurately quantifying these WMHs is crucial for understanding their clinical impact, but current manual segmentation methods are time-consuming, laborious, and prone to subjectivity. This leads to a need for automated, objective, and reliable tools for WMH segmentation.

Methods

The researchers developed a novel U-shaped 2D convolutional neural network (CNN) architecture for automatic WMH segmentation. The model was trained on a large dataset of manually annotated FLAIR MRI scans. Key aspects of their approach include:

  • U-Net Architecture: A modified U-Net architecture was employed, known for its effectiveness in image segmentation tasks.
  • 2D Slice-based Processing: The model processes 2D axial slices of FLAIR images, with the final 3D mask generated by concatenating the predictions from individual slices.
  • Preprocessing: Standard preprocessing steps like global intensity scaling, zero-padding to square images, and resizing/cropping to a uniform 384x384 pixel dimension were applied.
  • Data Augmentation: Techniques such as vertical/horizontal flips, rotation, contrast enhancement, scaling, translation, and Gaussian noise were used to improve model generalizability.
  • Evaluation Metrics: Performance was assessed using per-pixel metrics (precision, recall, dice score, accuracy) and lesion-wise metrics to account for the importance of correctly identifying individual lesions, especially smaller ones.

Datasets

  • Development Set: 996 patients' FLAIR MRI data from Beijing Tiantan Hospital (September 2018 - September 2019) were used for training and validation.
  • Internal Test Set: 160 patients' data from Beijing Tiantan Hospital were used for independent internal testing.
  • External Test Set: 90 patients' FLAIR MRI data from three other hospitals across China (part of the Third China National Stroke Registry - CNSR-III) were used for multi-center external validation.

Key Insights and Ideas

  • Clinical Utility is Paramount: Beyond just achieving high quantitative scores, the paper emphasizes the importance of clinical evaluation. The high acceptance rate by neuroradiologists suggests the system is practical for real-world use.
  • Lesion-wise Evaluation Matters: The authors highlight that per-pixel metrics can be biased towards larger lesions. Incorporating lesion-wise evaluation provides a more comprehensive understanding of the model's ability to detect and delineate individual WMH lesions, regardless of size.
  • Preprocessing is Crucial: The study notes that preprocessing steps, including image scaling and resizing, play a significant role in the DLS development and performance.
  • Generalizability Across Centers: The system demonstrated consistent performance on both internal and external test sets, indicating good generalizability across different data sources and scanners.

Key Results

  • High Accuracy: The proposed CNN architecture achieved an average pixel-wise dice score of 0.711 and a lesion-wise dice score of 0.735 on the external test set.
  • Superior Performance: The system outperformed state-of-the-art deep learning architectures like U-Resnet and 3D-unet in head-to-head comparisons.
  • Clinical Acceptance: In a qualitative clinical evaluation, 95% of the segmentation results were deemed acceptable or acceptable with minor revision by three independent neuroradiologists.
  • Inter-rater Agreement: The Kendall W test indicated good inter-rater agreement among the neuroradiologists, suggesting a consistent standard for clinical assessment.

Limitations and Future Work

  • Retrospective Nature: The study was retrospective, and the data came from multiple scanners. While this can enhance robustness, more extensive testing on prospective studies is needed.
  • Single Modality: The DLS currently relies solely on FLAIR MRI. Future work could explore incorporating information from other MRI sequences to potentially improve segmentation accuracy.
  • Data Source Variation: While the system showed generalizability, the authors acknowledge that the analysis did not exhaustively include all scanner types and acquisition protocols. Further testing across a wider range of data sources is recommended.

Summary updated

Abstract

<b>Objective:</b> Reliable quantification of white matter hyperintensities (WHMs) resulting from cerebral small vessel diseases (CSVD) is essential for understanding their clinical impact. We aim to develop and clinically validate a deep learning system for automatic segmentation of CSVD-WMH from fluid-attenuated inversion recovery (FLAIR) imaging using large multicenter data. <b>Method:</b> A FLAIR imaging dataset of 1,156 patients diagnosed with CSVD associated WMH (median age, 54 years; 653 males) obtained between September 2018 and September 2019 from Beijing Tiantan Hospital was retrospectively analyzed in this study. Locations of CSVD-WMH on the FLAIR scans were manually marked by two experienced neurologists. Using the manually labeled data of 996 patients (development set), a U-shaped novel 2D convolutional neural network (CNN) architecture was trained for automatic segmentation of CSVD-WMH. The segmentation performance of the network was evaluated with per pixel and lesion level dice scores using an independent internal test set (<i>n</i> = 160) and a multi-center external test set (<i>n</i> = 90, three medical centers). The clinical suitability of the segmentation results, classified as acceptable, acceptable with minor revision, acceptable with major revision, and not acceptable, was analyzed by three independent neuroradiologists. The inter-neuroradiologists agreement rate was assessed by the Kendall-W test. <b>Results:</b> On the internal and external test sets, the proposed CNN architecture achieved per pixel and lesion level dice scores of 0.72 (external test set), and they were significantly better than the state-of-the-art deep learning architectures proposed for WMH segmentation. In the clinical evaluation, neuroradiologists observed the segmentation results for 95% of the patients were acceptable or acceptable with a minor revision. <b>Conclusions:</b> A deep learning system can be used for automated, objective, and clinically meaningful segmentation of CSVD-WMH with high accuracy.
BibTeX
@inproceedings{shan2021segmentation,
  author = {Wei Shan and Yunyun Duan and Yu Zheng and Zhenhua Wu and Wei Sheng Chan and Qun Wang and Peiyi Gao and Yaou Liu and Kunlun He and Yongjun Wang},
  title = {{Segmentation of Cerebral Small Vessel Diseases-White Matter Hyperintensities Based on a Deep Learning System}},
  booktitle = {Frontiers in Medicine},
  year = {2021},
  doi = {10.3389/fmed.2021.681183},
}
Figure