This is a retrospective study and was approved by the institutional review board in accordance with local ethics procedure.
Study Population and participants
Imaging Protocol
CT scans were performed using a third-generation 256-slice dual source computed tomography system (AQUILON Prime-SP, Canon Medical Systems, SOMATOM Definition Flash, Siemens Healthcare, Forchheim, Germany). All scans were performed in a craniocaudal direction with a standard prospective cardiac-gated protocol [23]. Exposure interval was chosen depending on the heart rate (HR): 30–80 % RR interval for patients with an HR ≥ 70 bpm, 70–80 % RR interval for patients with an HR < 70 bpm. Acquisition range covered 1 cm below the carina to cardiac apex. Scanning parameters were defined as follows: slice collimation of 2 × 64 × 0.6 mm, the field of view of 220 × 220 mm, gantry rotation time of 280 ms, tube voltage of 120 kV, and the tube current was adjusted as a function of patient size. CT images were reconstructed at axial, sagittal, and coronal
Ground Truth Assessment
CAC identification and isolation
A Cardiologist with Level 2 CT accreditation (S.R., with Ten years of experience in cardiovascular CT image and reported > 5000 studies) manually performed CAC identification and isolation on an axial and sagittal view of the CT images. CAC was identified using 130Hu as a threshold in the coronary artery distribution and isolated [24].
CAC calculation and evaluation
Agatston score (AS), was estimated based on manual identification and isolation of coronary calcification, which were obtained by proprietary calcium scoring software according to clinical calcium scoring standards [27]. The result of CACS was respectively recorded at total and vessel-specific levels, i.e., left main artery (LM), left ascending artery (LAD), left circumflex artery (LCX) and right coronary artery (RCA).
Risk Categorization
At the patient’s level, according to CAC-DRS recommendation [7], CACS were recorded as 5 grades with total AS result: Category I (Agatston score 0, A0), Category II (Agatston score 1-10, A1), Category III (Agatston score 11-100, A1), Category IV (Agatston score 101-400 A2,), Category V (Agatston score (> 400, A3). The number of calcified vessels was recorded as N0–N4 for each patient.
Deep Learning Supported Model
We have developed a novel approach for Calcium detection, classification, and quantification in a totally automatic fashion from the given CT volume. Main steps in building this automated solution are namely (Figure 1):
- Multi Organ Segmentation: Cardiac, Aorta, and epicardial Fat segmentation.
- Coronary artery region segmentation.
- True calcium segmentation.
- Calcium Scoring.
The input cardiac CTs are uploaded using a software as a service (SAS) based platform developed by our team. The data was anonymized to remove patient health information (PHI) information from the scans. The models were trained on eight NVIDIA Tesla V100 GPUs running on AWS cloud infrastructure (Fig 1). The input to the model is a 3D non-contrast cardiac CT scan and images are consistent with 512 × 512 format. The images were resized to match the input size. This is followed by morphological dilation and anisotropic smoothing. Next, we segment the calcium in the heart by first using intensity and volume threshold to filter calcium in the non-contrast CT and then removing noise as well as non-useful calcium in the bone and extra cardiac calcium. The Agatston score for each artery region is then computed as the sum of Agatston scores of calcium group in that artery region across all slices.
Multi Organ Segmentation: Cardiac, Aorta, and epicardial Fat segmentation.
We used our in-house modified state-of-art densely connected deep network framework KardioNet, which has been customized to multi organ segmentation of the cardiac scan. The proposed method is based on the existing well-known multi-instance segmentation method, Mask R-CNN (25). The architecture consists of two main stages (Figure 2). The first stage of the region proposal network comprises of identifying the object bounding boxes Next, the highest-ranking bounding boxes are identified and used to generate region proposals, thus focusing algorithm attention on specific regions of the image.
The mask generated by the segmentation model had some noise that included non-relevant subbranches of the main artery and noise due to other cardiac structures. We applied post processing step to further improve the segmentation mask. The proposed framework segments the region of interest and identifies the calcium by applying localization to segment different voxel regions. To achieve this, we branched two instances of the architecture, the first runs a segmentation subnetwork for identifying heart and aorta region and the second runs a subnetwork which identifies epicardial fat and thus is integrated to the above segmentation to improve the precision of multiorgan localization. To segment individual arteries, we use the fact that epicardial fat surrounds the arteries and has a particular density range, which allows us to obtain an epicardial fat mask and hence to segment the arteries. Next, we segment the calcium in the heart by first using intensity threshold to filter calcium in the non-contrast CT and then removing noise as well as non-useful calcium in the bone and outside the heart. Applying the coronary artery masks to the segmented calcium then gives us calcium segments in each artery. The Agatston score for each artery is then computed as the sum of Agatston scores of calcium group in that artery across all slices.
Cardiac and aorta segmentation
We performed cardiac and aorta segmentation. Experienced radiologist generated masks were used to train the model (40 cases with 40-50 slices in each mask were used to train the model). Aorta segmentation was performed on the axial slices. Axially the shape of the aorta changes from circular to elliptical to semi-elliptical before it disappears. We use these shape changes in the prediction for the mask, which yields a segmentation (Mean Dice score :0.91). Figures 3a, and 3b show sample heart and aorta segments, respectively.
Coronary artery region segmentation using epicardial fat
The coronary arteries lie in and are surrounded by the epicardial fat on the epicardial surface; therefore, we locate them in each region with the help of epicardial fat mask. Epicardial fat has a fixed range of density, which allows us to derive a mask for it from the heart mask. Figure 3c shows a sample epicardial fat segment. Coronary artery regions are then located using dilation and erosion on the derived epicardial fat mask. Finally, heart region splitting is used to locate individual coronary artery regions.
Heart region splitting and coronary artery segmentation
As CAC quantification needs to be specific to the type of artery, we divide the heart into regions containing the corresponding artery. Iterating from the top, the first slice containing elliptical aorta is chosen as a reference and a slice a few millimetres below is used as localization for splitting the heart into various regions. In the remaining slices, we segment the aorta but retain the artery regions, which divides the slices into aorta, right coronary artery (RCA), left anterior descending (LAD), and left circumflex (LCX) regions.
Coronary artery and aorta calcium segmentation
For CAC scoring in coronary arteries and ascending aorta, we need to segment the relevant calcium. To filter calcium, we use an intensity threshold (intensity > 130 HU) in non-contrast CT. This filtered calcium also includes non-relevant calcium (noise, artifacts, and bones). We eliminate most of the noise using 2D erosion and dilation and the rest using a volume and maximum intensity threshold. Next, to eliminate bone calcium, we apply 3D dilation only on calcium groups with volume greater than 10,000 voxels (3000 mm3) to merge the tiny bone calcium groups into larger ones and eliminate them together. In the above process, some parts of relevant calcium groups may have been removed; we use connected components technique to recover those parts. Finally, we apply the heart mask to segmented calcium to eliminate extra cardiac calcium groups. Figure 3d, and 3e shows a sample calcium segment in the coronary arteries. On this segmented cardiac calcium, we apply aorta and individual coronary artery (RCA, LAD, and LCX) masks to obtain calcium segment in each artery. Once again, we use connected components technique to recover relevant calcium that may have been lost.
Computing Agatston score
The Agatston score for each of the aorta and coronary arteries is obtained by summing the Agatston scores for the calcium groups in the artery in each slice and summing these scores over all slices. The Agatston score for an individual calcium group (lesion) is computed by multiplying the lesion area with the corresponding coefficient in Table 1, which assigns a coefficient to the intensity range in which the point of maximum intensity in a lesion lies.