TL;DR
This paper introduces a deep learning model that can predict cardiovascular disease (CVD) risk directly from low-dose computed tomography (LDCT) scans, typically used for lung cancer screening. This approach allows for simultaneous CVD risk assessment in high-risk individuals, potentially reducing the need for separate, more resource-intensive cardiac imaging.
What Problem Does This Paper Address?
Cardiovascular disease (CVD) is a leading cause of death, and individuals with cancer have a significantly higher risk of CVD mortality. Lung cancer screening programs, which increasingly utilize low-dose computed tomography (LDCT), present a unique opportunity to assess CVD risk in a population already at higher risk for both conditions. However, current standard practice for CVD risk assessment often requires dedicated cardiac CT scans, which involve higher radiation doses and costs. This paper addresses the challenge of extracting meaningful CVD risk information from existing LDCT scans, which are not optimized for cardiac imaging and can suffer from artifacts and lower signal-to-noise ratios.
Methods
The researchers developed a two-stage deep learning approach:
- Heart Detector: A RetinaNet-based convolutional neural network (CNN) was trained to accurately identify and isolate the heart region within the large field of view of LDCT scans. This focuses the subsequent analysis on the most relevant anatomical area.
- CVD Risk Prediction Model (Tri2D-Net): This is a 3D CNN designed to process the heart region. It utilizes three separate 2D CNN branches, each processing the heart region from axial, sagittal, and coronal views. This "Tri2D" approach aims to efficiently extract features without the computational burden of a full 3D CNN. An attention mechanism was incorporated to help the model focus on features known to be correlated with CVD, such as calcification and epicardial/pericardial fat. The model was trained to classify subjects as CVD-positive or CVD-negative, with the output probability serving as a CVD mortality risk score.
Datasets
- National Lung Screening Trial (NLST): This public dataset, comprising LDCT scans from 10,395 subjects, was used for model development and internal validation. It includes information on CVD-related deaths.
- Massachusetts General Hospital (MGH) Dataset: This independent dataset, consisting of LDCT scans from 335 subjects, was used for external validation. Crucially, this dataset also included ECG-gated cardiac CT (CCT) scans, allowing for comparison with established CVD risk markers like Coronary Artery Calcification (CAC) score, CAD-RADS score, and MESA 10-year risk score.
Key Insights and Ideas
- LDCT as a Dual-Purpose Tool: The study demonstrates that LDCT scans, primarily intended for lung cancer screening, contain sufficient information to accurately assess CVD risk. This opens the door for a more integrated screening approach.
- Beyond Calcification: The model's ability to incorporate features beyond just coronary artery calcification (e.g., juxtacardiac fat) suggests a more comprehensive understanding of CVD risk from LDCT, potentially capturing subclinical markers missed by traditional methods.
- Reducing Inter-Observer Variability: By providing an automated, quantitative CVD risk score, the deep learning model has the potential to reduce the significant variability observed among human expert interpretations of cardiac imaging.
- Potential for Cost and Radiation Reduction: By leveraging existing LDCT scans, this approach could reduce the need for additional cardiac CT scans, thereby lowering healthcare costs and patient radiation exposure.
Key Results
- NLST Dataset: The deep learning model achieved an Area Under the Curve (AUC) of 0.871 for CVD screening and 0.768 for CVD mortality risk prediction on the NLST test set. It significantly outperformed other reported deep learning methods.
- MGH Dataset Validation: On the independent MGH dataset, the model achieved an AUC of 0.924 for CVD screening. When compared against gold-standard CVD risk scores:
- CAC Score: AUC of 0.881
- CAD-RADS Score: AUC of 0.763
- MESA 10-year Risk Score: AUC of 0.835 In all comparisons, the model either outperformed or showed strong consistency with these established markers.
Limitations and Future Work
- ICD-10 Coding Limitations: The NLST dataset relied on ICD-10 codes for labeling CVD-related deaths, which may lead to misclassification or missed CVD deaths. The MGH dataset's use of gold-standard CVD risk scores helped mitigate this.
- Limited Mortality Data: Access to sufficient mortality information from LDCT scans is still developing, as widespread LDCT screening for lung cancer is relatively recent.
- Generalizability to Diverse Scanners: While validated on an independent dataset, further testing across a wider range of scanner manufacturers and protocols would strengthen generalizability.
- Future Work: The authors suggest that this approach could be extended to other clinical scenarios, such as predicting mortality in pediatric cancer patients. They also highlight the potential for this technology to reduce healthcare disparities by providing accessible CVD risk assessment.
Summary updated