Abstract
Purpose: Image fusion strategies of myocardial perfusion imaging (MPI) and coronary CT angiography (CCTA) have shown increased diagnostic power. However, their clinical feasibility is hindered by the lack of efficient algorithms for the extraction of cardiac anatomy from CCTA datasets. The aim of this work was to validate our previously published algorithm for automated cardiac segmentation of CCTAs in a larger cohort of subjects while testing its application in clinical settings. Methods: Three borders were automatically and manually extracted on sixty-three clinical CCTAs: left and right endocardia (LV, RV) and the biventricular epicardium (EPI). Impact of image resolutions and inter-operator variability on accuracy and robustness of automated processing were evaluated. Automated algorithm accuracy was assessed with the Dice Similarity Coefficient (DSC) and the surface-to-surface distance metric. Relevant quantities were compared for automated versus manual segmentations: LV and RV volumes, myocardial mass and LV myocardial mass. Results: Lower resolution images offered an acceptable trade-off for accuracy and processing time (45 sec). DSC for LV, RV, EPI borders were 0.88, 0.80 and 0.89. Automated versus manual correlation coefficients for LV and RV vol, myo and LV mass were 0.96, 0.73, 0.84 and 0.67 with inter-operator agreement > 0.93 for three variables. Consistent and improved results were evidenced at higher resolutions. Conclusion: Our algorithms allowed efficient automated cardiac segmentation from CT imagery with minimal user intervention, clinically acceptable times and accuracy. The reported results show promise for its use in a clinical environment, specifically in the context of image fusion.
Introduction
To diagnose Coronary Artery Disease (CAD), clinicians must evaluate both cardiac anatomy (via Coronary Computed Tomographic Angiography, or CCTA) and physiology (via Myocardial Perfusion Imaging, or MPI). The "holy grail" of cardiac diagnostics is multimodality image fusion—overlaying these datasets to correlate specific arterial blockages with blood flow deficits.
However, a significant bottleneck exists: aligning these images requires "segmenting" or digitally tracing the heart's borders in the CCTA scan. Manually performing this task is a grueling 2-to-3-hour process per patient, making it impractical for clinical use. This study validates an automated approach designed to reduce this processing time to under a minute.
Methods and Materials
The study utilized 63 clinical CCTA datasets collected from international sites and various scanner manufacturers to ensure the algorithm's robustness across different clinical environments. The automated approach employs a modified Chan-Vese technique, a type of "active contour" algorithm that evolves boundaries for the left ventricle (LV), right ventricle (RV), and epicardium (EPI) simultaneously.
To maintain anatomical realism, the algorithm incorporates two key constraints:
- Statistical Shape Priors: Principal Component Analysis (PCA) was used to teach the algorithm the expected shape of a human heart, preventing it from wandering into surrounding tissues.
- Coupled Evolution: The LV and RV boundaries are evolved together with a penalty for overlapping, ensuring the heart walls maintain realistic thickness.
Key Results and Insights
The automated process demonstrated high efficiency and clinically acceptable accuracy:
- Speed: The algorithm processed low-resolution images in just 45 seconds, compared to hours for manual segmentation.
- Accuracy: The Dice Similarity Coefficient (DSC) reached 0.88 for the LV and 0.80 for the RV. Correlation for LV volumes was excellent (R = 0.96).
- Resolution Trade-offs: Researchers found that processing lower-resolution images (128³ voxels) provided a significant speed boost with only a minor impact on accuracy, suggesting a viable path for real-time clinical application.
- User Consistency: The tool requires minimal user input (defining intensity points for bone and background), resulting in high inter-operator agreement regardless of the user's experience level.
One persistent challenge noted was the segmentation of the Right Ventricle. The RV wall is thinner and often exhibits lower contrast than the LV, making it a more difficult target for automated tools and human experts alike.
Conclusion
This validation study demonstrates that automated biventricular myocardial segmentation can achieve high accuracy within timeframes suitable for a busy clinical environment. While future work will focus on improving segmentation at the heart's base and apex, the current results show great promise for integrating this tool into "one-click" software for CCTA and MPI fusion, potentially making advanced multimodality imaging a standard-of-care diagnostic tool.
Summary updated