Vessel segmentation for X-ray coronary angiography using ensemble methods with deep learning and filter-based features

Zijun Gao, Lu Wang, Reza Soroushmehr, Alexander Wood, Jonathan Gryak, Brahmajee K. Nallamothu, Kayvan NajarianBMC Medical Imaging 2022 · DOI

TL;DR

This paper introduces a novel ensemble framework for segmenting coronary arteries in X-ray coronary angiography (XCA) images. By combining traditional filter-based features with deep learning features, their method achieves superior performance compared to existing deep learning models, offering a more robust and consistent approach for CAD diagnosis.

What Problem Does This Paper Address?

Coronary artery disease (CAD) is a leading cause of death globally, and accurate assessment of coronary artery stenosis is crucial for diagnosis and treatment. X-ray coronary angiography (XCA) is the gold standard for this, but segmenting the coronary arteries in these images is challenging due to low signal-to-noise ratios, low contrast, confounding background structures (like catheters and spine), and the complex 3D projection into 2D images leading to twisted and overlapping vessels. Existing automated segmentation methods, while improving, still struggle with these challenges, leading to unreliable stenosis assessments and potentially impacting patient care.

Methods

The core of this work is a novel ensemble framework that leverages both traditional computer vision techniques and deep learning. The process involves:

  • Feature Extraction:
    • Filter-based Features: A comprehensive set of 21 features were extracted using various filters (e.g., Frangi, matched, Gabor filters) and scale-space theory. These capture morphological, spatial, and frequency domain characteristics of the vessels.
  • Deep Learning Features: Features were extracted from the activation maps of the final decoder layer of a trained DenseNet-backbone U-Net model. This provides high-level, contextual information.
  • Combined Feature Vector: These two sets of features were concatenated to create a rich 37-dimensional feature vector for each pixel.
  • Ensemble Classifiers: Two powerful ensemble learning methods were employed:
  • Gradient Boosting Decision Tree (GBDT): A sequential boosting algorithm that iteratively builds a strong model from weak learners.
  • Deep Forest: A deep ensemble model that uses non-differentiable modules and cascade structures for feature transformation.
  • Under-sampling: To address the severe class imbalance (vessel pixels vs. background pixels), a hybrid under-sampling strategy was used. This involved:
  • Uniform and Unsupervised Under-sampling: To reduce the majority (background) class.
  • Tomek Links: To remove overlapping pixels between classes and create clearer boundaries.
  • Cluster Centroid: A more aggressive method to balance classes by replacing sample points with cluster centroids.

Datasets

The study utilized a dataset of 130 XCA images from two sources: the University of Michigan Hospital and a hospital in the United Kingdom. These images were de-identified and included patients suspected of CAD. The dataset was split into training, validation, and test sets in a 3:1:1 ratio. No new benchmark datasets were introduced; the focus was on applying novel methods to existing data.

Key Insights and Ideas

  • Hybrid Feature Engineering is Powerful: Combining handcrafted filter-based features with learned deep features provides a more comprehensive representation of vessel structures than either approach alone. This is particularly beneficial when dealing with complex image characteristics like those in XCA.
  • Ensemble Methods Offer Robustness: Ensemble methods, especially GBDT, demonstrated superior performance and consistency compared to individual deep learning models. This suggests that ensemble approaches can better handle the inherent noise and variability in XCA images, especially with limited training data.
  • Under-sampling is Crucial for Imbalanced Data: The severe imbalance between vessel and background pixels necessitates careful under-sampling. The study showed that different under-sampling strategies (especially Cluster Centroid) significantly impacted the performance of the ensemble models, highlighting its importance.
  • Deep Forest Shows Promise: While GBDT achieved slightly better results, Deep Forest also outperformed standard deep learning models, indicating its potential for medical image segmentation tasks.

Key Results

The best performing model, a GBDT classifier trained on data processed with the Cluster Centroid under-sampling method, achieved:

  • Mean Precision: 0.857
  • Mean Sensitivity: 0.902
  • Mean Specificity: 0.992
  • Mean F1 Score: 0.874
  • Mean AUROC: 0.947
  • Mean IoU: 0.779

These results outperformed several common deep learning models (U-Net variants, DeepLabV3+) and a state-of-the-art method [43] across most metrics, with lower standard deviations, indicating more consistent performance.

Limitations and Future Work

The authors acknowledge several limitations:

  • Computational Cost of Cluster Centroid: The Cluster Centroid under-sampling method is computationally intensive, significantly increasing training time.
  • Feature Importance Analysis: While permutation importance was used, it can be biased by correlated features. A more substantial analysis is needed to identify truly discriminating features.
  • Generalizability of Deep Features: The deep neural networks used for feature extraction were trained with a simplified logic (early stopping) to avoid overfitting on the small dataset. While this was deemed necessary, it might lead to slightly inferior deep features compared to models trained for longer.
  • Transformer Networks: The paper notes that while vision transformers show promise, they are currently impractical for medical image segmentation with limited data due to high computational demands and the need for extensive training data.

Future work could focus on:

  • Developing more efficient under-sampling techniques.
  • Conducting a deeper analysis of feature importance to build more explainable models.
  • Exploring alternative deep learning architectures for feature extraction that balance performance and generalization on smaller datasets.
  • Investigating transformer-based models if computational resources and data availability improve.

Summary updated

BibTeX
@article{gao2022vessel,
  author = {Zijun Gao and Lu Wang and Reza Soroushmehr and Alexander Wood and Jonathan Gryak and Brahmajee K. Nallamothu and Kayvan Najarian},
  title = {{Vessel segmentation for X-ray coronary angiography using ensemble methods with deep learning and filter-based features}},
  journal = {BMC Medical Imaging},
  year = {2022},
  doi = {10.1186/s12880-022-00734-4},
}
Figure