Convolution Neural Network-Based Automatic Scoring of Coronary Artery Calcification and Its Value to Identify the Risk of Radiation-Induced Cardiac Toxicity on Radiotherapy Planning CT Scans in Breast Cancer Patients

Abstract

To develop an automatic scoring of coronary artery calcification (CAC) on breast cancer radiotherapy (RT) planning computed tomography (CT) scans, and to explore its predictive value of CAC for radiation-induced cardiac toxicity. Planning CT scans of 668 breast cancer patients from two prospective clinical trials (NCT02942615, NCT03829553) were retrospectively reviewed. In total, 34 CTs containing CAC were identified. The training and test samples were 29 and 5, respectively. We proposed a two-stage model for CAC segmentation task with nnU-Net as backbone. The segmentation results were processed by threshold extraction and region growth algorithm. We also employed transfer learning to automatically identify calcification of left anterior descending artery (LAD), right coronary artery (RCA), left circumflex artery (LCX), and left main coronary artery (LM) based on a public dataset of 430 cases from Stanford University. The data of cardiac examination of these 34 patients before and during the follow-up after RT were collected. The cardiac event was any symptomatic heart disease or new-onset abnormality in the cardiac examination after RT. The mean dice coefficients (DSC) and 95% Harsdorf distance (95HD) of test samples were 0.992 and 0.599 mm, respectively. The mean absolute error (MAE) of CAC Angaston score between ground truth (GT) and predictions was 0.532. The detailed consistency parameters of 5 test samples were shown in Table 1. After 1:2 propensity score matching (PSM), 21 patients had CAC and 42 patients had no CAC were selected. The number of patients with CAC scores of 1 to 10, 11 to 100, and greater than 100 was 10, 9 and 2, respectively. During median follow-up of 9.2 months (range, 1-42.7), 90.5% and 38.1% of patients in CAC cohort and no CAC cohort developed cardiac event (p<0.001). Patients with CAC had significantly increased cardiac events (HR = 2.4; 95% CI, 1.22-4.75; p = 0.0117). The risk of cardiac events increased with CAC scores ([HR]1-10 = 2.1, 95% CI 0.9-4.9; [HR]11-100 = 2.5, 95% CI 1.0-5.9; [HR]>100 = 4.0, 95% CI 0.9-17.4). Our primary results showed that this two-stage segmentation model is capable of achieving automatic CAC scoring which might assist to predict the risk of post-RT cardiac events in breast cancer patients.