An Optimized Semantic Segmentation Deep Learning Model for Automated Coronary Artery Calcification Prediction
作者
Nathan K. Yee
出处
期刊: [Institute of Electrical and Electronics Engineers] 日期:2024-07-19卷期号:: 225-229被引量:3
标识
DOI:10.1109/prml62565.2024.10779669
摘要
Coronary artery disease (CAD) is the leading cause of death in the United States, accounting for 25% of all deaths. Early detection by Coronary Artery Calcium (CAC) scoring of chest computed tomography (CT) scans is crucial but remains a significant challenge due to high costs, time demands, and the need for specialized experts and equipment. While current AI-based methodologies are promising, significant hurdles have impeded their clinical adoption. This study is the first effort to benchmark image classification and image segmentation based deep learning models in direct comparison and demonstrates that image segmentation techniques outperform in automated CAC scoring when benchmarked with image classification. The image classification is performed using ResNet-50, and the image segmentation is performed with an architecture based on Attention U-Net. The study benchmarks the models based on the Stanford AIMI COCA database with gated computed tomography scans of 790 patients and nearly 40,000 total image files. Additionally, this study implements an image optimization strategy that utilizes contrast limited adaptive histogram equalization (CLAHE) and Gaussian blur specific to the CAC score range, thereby predicting a patient's risks for CAD and recommending the appropriate preventative action or treatment. This approach effectively enhances predictive accuracy to above 90% for most CAC score ranges and Dice Coefficient (DCE) to approximately 0.86, a substantial improvement compared to previous efforts using image segmentation deep learning. With increased reliability, the proposed deep learning system demonstrates the viability of AI-driven CAC monitoring for those at risk of CAD to decrease mortality rates.