This study was conducted according to the principles set forward in the declaration of Helsinki and according to Good Clinical Practice. Permission was obtained from the regional ethical review board (Dnr 2018/535-32). In accordance with the ethical regulations for Swedish registries and Swedish legislation, patients were informed about their participation in a registry, and the right to deny participation or have data removed, which waivers any requirements for written consent.
For the multi-hospital acquired training data, necessary ethics approval and/or patient consent was obtained when required.
In this observational, cross-sectional study, patients and their baseline characteristics were retrospectively collected from a nationwide quality registry, SWEDEHEART . All patients with a CSCT performed on a particular state-of-the-art CT scanner between 1 December 2017 and 31 January 2019 (n = 342) at Linköping University Hospital were consecutively included. The indication was suspected ischemic heart disease. Exclusion criteria were, as previously suggested , anatomical abnormalities (n = 2), intracoronary stents (n = 0), metal implants (n = 18), and CSCT scans with severe motion artifacts or high level of noise determined by visual inspection (n = 6) (Fig. 1). In addition, one CSCT scan (n = 1) was excluded due to incomplete scanning of the coronary arteries. The final main dataset consisted of 315 CSCT scans.
In total, 13 out of 20 CSCT with anatomical abnormalities and metal implants were considered readable for CAC scoring, and represented an independent dataset (n = 13).
CT acquisition parameters and image reconstruction
All CSCT scans were acquired through the use of a Siemens SOMATOM Force (Siemens Healthineers) MDCT. A prospectively ECG-triggered high-pitch spiral CSCT scan was performed, with a tube voltage of 120 kV, and automated tube current modulation (CARE Dose4D, Siemens) with a setting of 40 quality ref. mAs. Further settings were as follows: gantry rotation time 0.25 s, pitch 3.2, collimation 192 × 0.6 mm, matrix size 512 pixels, and temporal resolution 66 ms. The scan was set to start at 65% of the cardiac cycle. Reconstructions were made with a routine weighted filtered back projection (WFBP, Siemens) algorithm using medium sharp convolution kernel (Qr36), 3.0 mm section thickness, and increment 1.5 mm. Beta blockers were administered if the heart rate was > 65 bpm. After CTCS scanning, a CCTA was performed in the same session.
AI-based, automatic system overview
The automatic software was trained on multi-vendor, multi-scanner, and multi-hospital, anonymized data from routine coronary calcium scoring acquisitions. No training datasets were used in the current study.
During model training, the locations of the coronaries created a territory map in a heart-centric coordinate system. This map serves to assign prior likelihood of different voxels belonging to the coronary arteries.
For each evaluated CSCT scan, a model is used to segment the heart, to establish a heart-centric coordinate system. The pre-computed coronary territory weights are mapped to the local size and shape of the patient’s heart. All voxels > 130 HU are extracted. Around each voxel, an image patch is extracted to represent the local spatial characteristics, the prior likelihood from the territory map and the location (x, y, z) of the voxel in the heart-centric coordinate system. The model uses these features to make a prediction that this voxel belongs to the coronaries.
Some work [10,11,12] already used patient-specific, heart-centric coordinate systems, but relied on manually placed markers, or local image coordinates in combination with a computationally expensive registration to an atlas-based model [13, 14]. Another work used a heart segmentation but no further classification besides voxel intensity . As far as we know, the evaluated new machine learning model that combines the location within this coordinate system, the local image information around a voxel, and the coronary territory map is novel.
A standard reference was obtained with a semi-automatic, previously validated , post-processing software (syngo.via, Siemens Healthineers). All 315 CSCT scans were double read by two radiologists in at least two sessions (M.S. and S.S., both with 10 years’ experience of cardiac CT reading) and all interpretation differences were resolved by consensus. To determine the presence of CAC, an attenuation threshold was set at > 130 HU. Calcified coronary objects having an area of ⋝ 1 mm2 were included, as originally described  using default software settings. Every calcified region of interest was manually identified and marked to attain the total AS, VS, MS, and the number of calcified coronary lesions. The time used for the first read was registered.
A total of 62 (20%) CSCT scans from the standard reference underwent a second opinion evaluation from two additional readers (A.P., radiologist, 20 years’ experience of cardiac CT reading, and L.H., cardiac imaging radiographer, 2 years’ experience of cardiac CT research reading). CSCT scans selected for second opinion were those considered to potentially shift in risk category due to readers arbitrariness (n = 32), calcifications close to the coronary ostia (n = 27), or difficulties to discriminate peripheral calcified coronary lesions from noise (n = 3). After consensus was reached, two changes were made, both with AS difference ≤ 5, and there was no shift in risk category.
Based on the AS, CSCT scans were assigned into commonly used risk groups with cutoff values: CAC 0: No identifiable plaque, very low CV. CAC 1–10: Minimal plaque burden, low CV risk. CAC 11–100: Mild atherosclerotic plaque burden, moderate CV. CAC 101-400: At least moderate atherosclerotic plaque burden, moderately high CV risk. CAC > 400: Extensive atherosclerotic plaque burden, high CV risk .
For inter-reader agreement, a subset of 106 (33.6%) CSCT scans were randomly selected and assigned to two independent radiologists. One radiologist was assigned 71 CSCT scans (G.N., 1 year experience of cardiac CT reading) and one radiologist 35 CSCT scans (A.B., 16 years’ experience of cardiac CT reading), both blinded to previous results.
The automatic software was implemented in MeVisLab on a regular workstation. All CSCT scans (n = 315) were analyzed with the automatic software, retrieving the total AS, VS, MS, and number of calcified coronary lesions. The automatic system run-time and the time for a manual double-check of the results were registered. The double-check included a localization of all CAC, and to attain an image-based numerical correlation to the automatically derived number of calcified coronary lesion.
No human interaction was needed, except for loading data into the software. For each CSCT scan, a visual CSCT feedback with crosshairs was displayed in three dimensions, allowing multiplanar reconstructions (Fig. 2).
The readable CSCT scans (n = 13) with coronary abnormalities and metal implants were independently reported, following the same routine. However, another radiologist (G.W., 16 years’ experience of cardiac CT reading) performed the semi-automatic double-read, and there was no second opinion.
Continuous data are presented as mean ± standard deviation if normally distributed, or as median and interquartile range (IQR) if non-normally distributed. Categorical data are presented as numbers and percentages. Normality was tested with Shapiro-Wilk’s test. The correlation and agreement between the standard reference and the automatic software for the AS, VS, MS, and the number of lesions were assessed with Spearman’s rank correlation coefficient (⍴) and intraclass correlation coefficient (ICC), as appropriate for non-parametric data. Bland Altman plots displayed bias and limits of agreements within 95% confidence interval. Differences in risk classifications were assessed by weighed kappa analysis (κ) and accuracy. Inter-observer agreement was demonstrated with ICC and Spearman’s rank correlation coefficient (⍴). Difference in time was analyzed with Wilcoxon signed-rank test. A two-sided p < 0.05 was considered statistically significant. Randomization for inter-rater agreement was achieved by Excel (Microsoft Office 365); all other analyses were performed using IBM SPSS v.24 (IBM SPSS).