Framework for detection and localization of coronary non-calcified plaques in cardiac CTA using mean radial profiles

Muhammad Moazzam Jawaid, Atif Riaz, Ronak Rajani, Constantino Carlos Reyes-Aldasoro, Gregory G. SlabaughComputers in Biology and Medicine 2017 · DOI · OpenReview

TL;DR

This paper introduces a novel framework for automatically detecting and localizing non-calcified plaques in cardiac CT angiography (CTA) using mean radial profiles and a Support Vector Machine (SVM) classifier. This approach achieves a promising 88.4% detection accuracy and 83.2% localization accuracy, offering a significant step towards automated plaque quantification and improved diagnosis of coronary heart disease.

What Problem Does This Paper Address?

Coronary heart disease (CHD) is a leading cause of mortality globally, and early detection of coronary abnormalities is crucial for preventing severe cardiac events. While calcified plaques are easily identifiable in CTA due to their high intensity, non-calcified plaques (NCPs) are challenging to detect because their intensity is similar to surrounding tissues. Traditional invasive methods carry risks, and existing non-invasive techniques struggle with NCPs, especially those causing positive remodeling. This paper aims to address this challenge by developing an automated method for detecting and localizing these difficult-to-identify non-calcified plaques.

Methods

The proposed framework consists of several key stages:

  • Vessel Segmentation and Centerline Extraction: The coronary vasculature is first segmented using an improved Chan-Vese energy model, followed by sub-voxel thinning to extract precise vessel centrelines.
  • Mean Radial Profile Calculation: For each coronary segment, radial profiles are computed by averaging intensity within concentric rings. This technique is effective in capturing intensity abnormalities within the 3D tubular structure of the vessels.
  • Cylindrical Modeling: Coronary segments are approximated by curved cylindrical models to facilitate the analysis of radial profiles along the vessel's length.
  • Feature Extraction: Several features are extracted to represent the coronary segments, including:
  • Windowed Mean and Standard Deviation: Statistical measures of the inner four radial rings, capturing intensity variations along the segment.
  • Mid-Lumen Intensity (fmid): Intensity of the innermost ring, reflecting contrast medium concentration.
  • Mean Radius (frad): Encoding vessel remodeling impact.
  • SVM Classification: A Support Vector Machine (SVM) classifier is trained using these extracted features to differentiate between normal and plaque-affected coronary segments.
  • Plaque Localization: For segments classified as abnormal, a derivative-based analysis of the mean radial profile is employed to identify local extrema and pinpoint the start and end positions of the plaque.

Datasets

The study utilized three different datasets for training and evaluation:

  • Rotterdam CTA Dataset: 18 CTA volumes from publicly available sources, offering multi-platform data from different vendors.
  • Guy's and St. Thomas' Hospital Dataset: 12 CTA volumes acquired using a Siemens Somatom Definition scanner.
  • Semmelweis University Dataset: 2 CTA volumes from Budapest, Hungary.

Manual annotations from experts served as the ground truth for all datasets.

Key Insights and Ideas

  • Mean Radial Profiles as a Novel Feature: The paper highlights the effectiveness of mean radial profiles in capturing subtle intensity variations characteristic of non-calcified plaques, which are often missed by methods focusing solely on intensity dips.
  • SVM for Robust Classification: The use of an SVM classifier, trained on a comprehensive set of features, demonstrates strong performance in differentiating between normal and abnormal segments, even with the challenging nature of NCPs.
  • Derivative Analysis for Precise Localization: Employing first and second-order derivatives on the mean radial profile proves to be an effective strategy for accurately identifying the boundaries of plaques within abnormal segments.
  • Feature Engineering for Clinical Relevance: The careful selection and engineering of features like mid-lumen intensity and mean radius, informed by clinical understanding of plaque behavior and vessel remodeling, significantly improve classification accuracy.

Key Results

The proposed framework achieved the following key results:

  • Plaque Detection Accuracy: An overall accuracy of 88.4% was achieved against manual expert annotations, with a sensitivity of 93% and specificity of 80%.
  • Plaque Localization Accuracy: The plaque localization accuracy, measured by the Dice similarity coefficient, was 83.2%.
  • Reproducibility: The model demonstrated consistent performance across three different CTA datasets, indicating its robustness and reproducibility for generic CTA data.
  • Feature Impact: The addition of mid-lumen intensity and radius-based features improved classifier accuracy by approximately 11%.

Limitations and Future Work

The authors acknowledge the following limitations:

  • Under-estimation in Fully Occluded Segments: The method may under-estimate long plaques in fully occluded segments.
  • Low Detection Rate in Minor Segments: Detection rates are lower in minor coronary segments due to their reduced diameter.
  • Immature Plaque Detection: Detecting immature (mild to moderate) soft plaques is challenging due to uncertain intensity profiles.

Future work suggested includes:

  • Inter-observer Error Investigation: Incorporating manual input from multiple experts to assess classifier efficiency relative to inter-observer error.
  • Quantitative Analysis: Extending the work to perform quantitative analysis for localization accuracy and aggregate plaque volume in the coronary tree.
  • Deep Learning Integration: Replacing hand-crafted features with a convolutional neural network (CNN) to leverage machine learning advancements, though this would require a larger CTA data repository.

Summary updated

BibTeX
@article{jawaid2017framework,
  author = {Muhammad Moazzam Jawaid and Atif Riaz and Ronak Rajani and Constantino Carlos Reyes-Aldasoro and Gregory G. Slabaugh},
  title = {{Framework for detection and localization of coronary non-calcified plaques in cardiac CTA using mean radial profiles}},
  journal = {Computers in Biology and Medicine},
  year = {2017},
  doi = {10.1016/j.compbiomed.2017.07.021},
}
Figure