TL;DR
This paper introduces a novel deep learning framework, MVSC, for fully automated analysis of coronary artery calcium (CAC) scores from non-contrast cardiac CT scans. The framework achieves high accuracy in quantifying total and vessel-specific CAC, including the number of calcified vessels, offering a significant step towards more efficient and comprehensive cardiovascular risk assessment.
What Problem Does This Paper Address?
Coronary artery calcium (CAC) scores are crucial indicators for predicting major adverse cardiovascular events. Current clinical practice relies heavily on manual analysis of CT scans to quantify these scores, which is time-consuming, tedious, and prone to inter-observer variability. While automated methods exist, they often struggle with accurately detecting calcifications near the aortic root and annulus, and many do not comprehensively report vessel-specific scores or the number of calcified vessels, as recommended by guidelines like CAC-DRS. This paper aims to bridge this gap by developing a fully automatic framework that addresses these limitations, providing a more complete and reliable assessment of CAC.
Methods
The core of this work is the MultiView Shape Constraint (MVSC) framework, a multi-task deep learning model. MVSC leverages multiview learning, mimicking how radiologists examine 3D CT scans by integrating information from axial, coronal, and sagittal views. It combines two key components:
- Calcification Detection and Segmentation: A deep learning model (RDN and DilatedUNet) is used to accurately detect and segment calcifications.
- Shape Constraint: A novel "Multi-view Shape Constraint" module is employed to refine the segmentation and improve accuracy, particularly in challenging areas like the aortic root. This module helps to disentangle features for comprehensive calcification analysis.
- Regression Model: A separate regression model then quantifies the CAC scores (Agatston, volume, and mass) based on the segmented calcifications.
This approach allows for both the extent (scores) and distribution (vessel-specific, number of calcified vessels) of calcification to be assessed automatically.
Datasets
- Training Data: 186 non-contrast cardiac-gated CT scans were used for model training.
- Testing Data: 46 non-contrast cardiac-gated CT scans were used for independent evaluation.
- Total Scans: 232 scans were included in the study after excluding cases with poor image quality, metal implants, poor breath-holding, coronary artery origin anomalies, or absence of calcification.
The study used retrospective data collected between October 2018 and March 2019.
Key Insights and Ideas
- Multiview Integration is Key: The success of MVSC highlights the importance of integrating information from multiple views (axial, sagittal, coronal) to overcome the limitations of single-view analysis, especially for challenging calcifications near anatomical boundaries.
- Shape Constraints Improve Accuracy: Explicitly incorporating shape constraints into the deep learning model helps to improve the precision of calcification segmentation, leading to more reliable score quantification.
- Comprehensive Reporting is Achievable: The framework demonstrates that it's possible to automatically quantify not just the total CAC score but also vessel-specific scores and the number of calcified vessels, aligning with current clinical reporting recommendations.
- Excluding Aortic Calcification: The framework effectively distinguishes coronary artery calcification from aortic calcification, a common challenge for automated systems.
Key Results
- High Accuracy in CAC Quantification: The MVSC framework showed no significant differences compared to manual quantification for Agatston score (automatic 535.3 vs. manual 542.0, P=0.993), calcium volume score (automatic 454.2 vs. manual 460.6, P=0.990), and calcium mass score (automatic 128.9 vs. manual 129.5, P=0.992) at the total level. Similar high agreement was observed at the vessel-specific level.
- Accurate Vessel Count: The number of calcified vessels was also accurately identified automatically, with no significant differences compared to manual counts for total vessels (automatic 107 vs. manual 102, P=0.125) and individual vessels (e.g., LM: 15 vs. 14, P=1.000).
- Excellent Consistency: High Inter-class correlation coefficients (ICCs) were reported for all metrics (AS, VS, MS, and #CV) at both total and vessel-specific levels, indicating strong agreement between automatic and manual results.
- Fast Processing: The automated framework is highly efficient, with a computational time of approximately 1.5 seconds per 3D CT scan during the testing phase.
Limitations and Future Work
- Exclusion of Other Findings: The current framework is designed solely for CACS analysis. Future work should aim to incorporate the detection and reporting of other cardiac and non-cardiac findings, including aortic calcification, as recommended by CAC-DRS.
- Coronary Artery Origin Anomalies: Patients with coronary artery origin anomalies were excluded. Accurate segmentation in such cases requires more specialized training data.
- Age Differences in Datasets: The testing dataset had a significantly older patient population than the training dataset. While the impact was mitigated by evaluating on similar data types, further validation across diverse age groups is warranted.
- Single-Center, Single-Scanner Study: The study was conducted at a single center using a single CT machine. Future multi-center and multi-scanner studies are needed to validate the generalizability of the proposed framework.
- Segmental Analysis: While the framework quantifies vessel-specific scores, further work on detailed segmental analysis of calcification distribution is suggested for even finer risk stratification.
Summary updated