Performance of modified ResNet model for calcium score from non-contrast computed tomography

Abstract Aims Calcium score measured from non-contrast CT is considered as the reference method for risk assessment of coronary artery disease. We aimed to develop a modified ResNet layers model for calcium score assessment in a cohort of multi-ethnic Asian patients with suspected coronary artery disease. Methods This is a multicenter study which performs CT scans in 5000 Asian Admixture patients in Singapore from 3 different hospitals (APOLLO study). In this current study, 369 subjects (mean age 56 ± 11 years; 147 female) were analyzed. Of these patients, 36% were shown to have very low risk (CAC = 0), 49% had mildly increased risk (CAC = 1–99), 11% had moderately increased risk (CAC = 100–299), and 5% had moderate to severe risk (CAC>300). The number of labelled 2D axial calcium lesions were as follows: LAD – 929, LM – 101, LCX – 545 and RCA – 940. Data augmentation was carried out to increase the overall sample size and balance number of labelled calcium lesions in the dataset. The dataset was split at the patient level prior to data augmentation to prevent any data leakage issues (60% Training: 20% Validation and 20% Test). The calcium score model was developed by consisting of three modified ResNet layers (MRN), and a series of fully connected layers. Axial, coronal and sagittal CT images were fed into the modified ResNet individually and the output of the model was the probability of the calcium lesion being in any of the four arteries. Cross-entropy loss was used as the loss function during the training process, while F1 score was used to evaluate the model performance at per-voxel level. Results The correlation coefficient between our MRN model and clinical ground truth for the total calcium score was r = 0.90 (P < 0.0001). The MRN model performed well in detecting the location of calcium lesions belonging to any of the three arteries, with accuracies of 0.95 in LAD, 0.90 in LCX, and 0.80 in RCA. Additionally, the MRN model demonstrated good diagnostic accuracy, F1 score, and intra-class correlation coefficient (ICC) of 0.80, 0.79, and 0.92, respectively, in predicting the CAC risk categories (low, mildly increased risk, moderately increased risk, and moderately to severely increased risk). A confusion matrix for the CAC risk category between the expert reader and MRN is shown in Figure 1. Conclusion Our modified ResNet-based machine learning model attained excellent prediction of calcium score using non-contrast CT. Figure 1.Confusion matrix for CAC risk.