Our overall approach to determine the role of the automated CCTA segmentation on MBF values was to measure the accuracy of MBFvs resulting from the fusion of dPET data with three differently extracted anatomical models. These models were as follows: (1) a totally manual (M) model, which also represented our selected standard of reference, (2) a fully automated segmented (FA) model, obtained with our published technique (Dahiya et al. 2019; Piccinelli et al. 2021), and (3) a corrected (C) model, where major inaccuracies in the automated segmentation were identified during quality control assessment of the borders and briefly edited by an expert in cardiovascular anatomy. Inaccuracies to be manually corrected were visually identified as automated contours that markedly deviated from the myocardial surface or the chambers edges. The resulting three sets of r/s MBFvs were compared to previously developed ranges of flow normality along major vessels in the three vascular territories (Piccinelli et al. 2020) and their agreement was measured with a technique based on DC.
Study population and imaging datasets
Multimodality image datasets collected in the context of the multisite DEMYSTIFY study (AlBadri et al. 2020) (ClinicalTrails.gov registration number NCT04221594) were used for this investigation. Each dataset is comprised of static, gated and dynamic r/s 13NH3 PET, CCTA and invasive coronary angiograph (ICA). The first 43 cases included in the DEMYSTIFY database were selected for this study. These patient data were contributed by collaborators in S. Korea: Seoul National University Hospital (SNUH, 35 cases), Samsung Medical Center (SMC, 2 cases) and Chonnam National University Hospital (CNUH, 6 cases). The image acquisitions were performed after local IRB committee approval and consent forms were obtained from all subjects.
All centers used equivalent clinical protocols for both cardiac PET and CCTA imaging as previously reported (Piccinelli et al. 2020; AlBadri et al. 2020). Briefly, prior to the cardiac PET, patients were asked to fast overnight, abstain from caffeine-containing beverages and stop vasodilator medications (such as beta-blockers or calcium channel blockers) for 24 h. The imaging protocol consisted of a low-dose CT scan for attenuation and scatter correction, followed by the injection of 370 MBq bolus of 13NH3 and the resting dynamic PET acquisition. Hyperemia was induced by a 3-min intravenous infusion of adenosine (140 μg/kg/min) followed by a second dose of 13NH3 and the stress PET acquisition. All images were acquired in list mode and binned according to the following temporal schemes: 12 × 10 s, 6 × 30 s, 2 × 60 s, 1 × 180 s for SNUH, 12 × 5 s, 6 × 10 s, 3 × 20 s, 6 × 30 s for SMC and CNUH. Prospectively ECG-gated contrast-enhanced CT images were acquired following standard clinical guidelines after the injection of 60 mL of nonionic contrast agent at 4 mL/s. Sublingual nitroglycerine was administered to all patients to facilitate visualization of coronary vessels. The diastolic phase (usually located between 65 and 85% of the cardiac cycle) was selected for the extraction of the anatomy as it allowed a relative motion-free visualization of the major vessels and the myocardium. Coronary CTA images were acquired either before or after cardiac PET. Image acquisition was performed after local IRB committee approval and consent forms were obtained from all subjects. All images were saved in DICOM format, anonymized and securely transferred to the core laboratory at Emory University for subsequent processing.
CCTA-derived anatomy extraction
A requirement of our approach for the calculation of MBF along coronary 3D paths is the extraction of the anatomical information from the CCTA images, specifically binary masks identifying LV, RV and EPI, which are used for the 3D image fusion of dPET and CCTA images, and coronary centerlines to localize the myocardium underneath the vessels. The extraction of the anatomical masks is a particularly challenging task. If manually performed, it can require up to two hours work by a trained user. Fully automated segmentations may require a quality control assessment, but they guarantee time efficiency and clinical feasibility. The technical details of our automated segmentation algorithms are presented elsewhere, as well as an initial validation of their robustness when compared to manual segmentation used as the reference standard (Dahiya et al. 2019; Piccinelli et al. 2021).
After re-orientation of the transaxial CCTA images in the short-axis (SA) direction, an isotropic volume of 512 × 512 × 512 pixels was created with the myocardium encompassing between 200 and 250 slices. A resampled version of this high-resolution dataset was created with a spacing 4 × the native planar one and used as input to the automated algorithms. M masks were obtained by manual delineation of the myocardium and blood pools on the high-resolution volume by an expert in cardiac imaging using an in-house developed software for editing myocardial borders. FA masks were generated with our proposed automated algorithms (Dahiya et al. 2019) from the low-resolution version of the SA-oriented CT images. Finally, FA masks were reviewed and C masks obtained as the result of a limited manual correction step. The manual corrections had the final goal of ensuring the completion of the fusion procedure and specific guidelines were provided to perform such corrections. Since the image registration (as described in the next section) relied on a univocal identification of the ventricles and myocardium, specific features of the segmented heart morphology (e.g., a clear separation between epicardium and endocardium along the septum and at the apex) were targeted. Only additional major segmentation errors (e.g., heavy RV encroachment into the liver space) were manually corrected, but minor ones ignored. The aim of this investigation was not to measure the performance of our automated segmentation tools, but to determine whether low-resolved automated or semi-automated segmented masks obtained in time-efficient fashion could guide the fusion phase in the context of our proposed method for MBFvs quantification and lesion significance assessment. Efforts were made to minimize user interactions in order to truly assess the MBFvs calculation when guided by FA masks and mimic clinical environment settings where brief high-level user guidance in automated medical image processing tasks is common. LV, RV and EPI borders were reviewed slice-by-slice and minimally corrected where/if necessary. The time devoted to these corrections was recorded for each case. Finally, coronary centerlines were manually extracted with in-house developed image processing tools (VMTK 2021) from CCTA images. Figure 1 shows the 3D biventricular models with coronary centerlines obtained from the 3 sets of anatomical masks.
PET/CCTA image fusion, flow model and vessel-specific MBF (MBFvs) calculation
As previously described (Piccinelli et al. 2020), our methodology to calculate flow along vessel centerlines relies on the fusion of dPET and anatomy. Importantly, the dPET sequence was first corrected for inter-frame motion (Nye et al. 2021) and the resulting images were used for the remainder of the processing. The registration procedure consisted of the following steps (performed for rest and stress separately): (1) a summed PET image (PETsum) was created from the dynamic frames of the second half of the acquisition interval; (2) the LVs segmented from the CCTA and PETsum were rigidly registered; and (3) mutual information-based techniques were used to align PETsum with the CCTA-derived biventricular mask using LV and RV structures from both modalities. More detailed information can be found in previous works (Faber et al. 2011; Piccinelli et al. 2018). The same transformations were finally applied to align the coronary centerlines to the PET space. Using the registered centerline as guidance, the myocardium subtended to each vessel was identified on the dPET images and discretized in contiguous cubic volumetric elements with a longitudinal size of 4 mm that followed the 3D trajectories. The MBF quantification was performed using standard tracer kinetic modeling techniques. Time-varying tracer concentrations were extracted in the form of time activity curves (TAC) from specific regions of interests (ROIs), namely vascular territories and the arterial blood, and used as input data into a 2-tissue compartmental model (Hutchins et al. 1990) that returned r/s MBF values for each identified ROI. The TACs from the contiguous cubic ROIs representing the myocardium subtended to vessels were also fed to the model, thus generating r/s flow profiles along the vessel (MBFvs). As the fusion procedure was performed with M, FA and C anatomical masks, 3 sets of r/s MBFvs curves were extracted for each analyzed vessel (respectively, \({\text{MBF}}_{{\text{M}}}^{{{\text{vs}}}}\), \({\text{MBF}}_{{{\text{FA}}}}^{{{\text{vs}}}}\) and \({\text{MBF}}_{{\text{C}}}^{{{\text{vs}}}}\)). Figure 2 shows the crucial processing steps for one of the analyzed cases: (A) the 3 sets of post-fusion coronary centerlines showing limited differences in orientation and position, (B) the extraction of the myocardium subtended to a vessel and its discretization in ROIs, and (C) the left anterior descending (LAD) r/s \({\text{MBF}}_{{\text{M}}}^{{{\text{vs}}}}\) presented as a two-dimensional (2D) graph of flow values along the vessel length (in mm, from the base to the apex of the heart). The actual flow values obtained from the cubic ROIs are indicated by the markers.
For the current investigation, r/s MBFvs profiles were extracted for one vessel per vascular territory for each analyzed case: LAD, one major vessel from the lateral wall (indicated as LCX), and the right posterior descending artery (rPDA). Figure 3A shows the anterior and posterior views of the heart with the vessel-specific modeling to probe three vessels for one of the analyzed cases. Figure 3B, C illustrates the differences introduced by the fusion of FA and C models with respect to the reference standard (M) in terms of LAD vessel-specific ROI spatial location.
Normal/abnormal classification of MBFvs curves
The MBFvs methodology was previously applied to define ranges of r/s MBFvs along the major vessels supplying the LV based on a cohort of patients (n = 15) with low risk of CAD who underwent 13NH3 PET and CCTA (Piccinelli et al. 2020). Since the final goal of this approach is to determine whether indexes based on MBF become abnormal at specific locations along a vessel, thus indicating the presence of flow-limiting lesions that require revascularization, these low-risk ranges were used to prospectively stratify MBFvs into normal (within the range) versus abnormal (below the range). In the context of this investigation, the low-risk ranges were used to determine whether the three sets of r/s MBFvs curves were classified in the same way and/or to what extent the automated and the corrected curves differed from the ones obtained when the manually delineated anatomy was used for image registration. Figure 4 displays 2D plots of the low-risk ranges and r/s \({\text{MBF}}_{{\text{M,FA,C}}}^{{{\text{vs}}}}\) for the three vessels considered for one of the analyzed cases.
To quantify the discrepancies in the normal/abnormal classification between \({\text{MBF}}_{{{\text{FA}}}}^{{{\text{vs}}}}\) and \({\text{MBF}}_{{\text{C}}}^{{{\text{vs}}}}\) compared to \({\text{MBF}}_{{\text{M}}}^{{{\text{vs}}}}\) curves, the Dice coefficient (Dice 1945) (DC) was calculated. DC measures spatial overlapping commonly used to rate the performance of segmentation algorithms with values between 0 (no overlapping) to 1 (complete overlapping). Adapted to our analysis DCFA was computed to quantify the number of flow values (i.e., ROI values) with matched classification between \({\text{MBF}}_{{{\text{FA}}}}^{{{\text{vs}}}}\) and \({\text{MBF}}_{{\text{M}}}^{{{\text{vs}}}}\) based on low-risk ranges. Analogously, DCC was derived from the comparison of flow values classification between \({\text{MBF}}_{{\text{C}}}^{{{\text{vs}}}}\) and \({\text{MBF}}_{{\text{M}}}^{{{\text{vs}}}}\). DCFA and DCC were computed globally by pooling all the ROI samples from all the vessels. A parallel analysis was conducted per vascular territory as image quantification performance may differ among vascular territories.
Inter-user variability analysis
To assess the impact of the user corrections of the anatomical masks on the fusion procedure and consequently on the MBFvs profiles, a second user was tasked with correcting the automatically extracted anatomy in a subset of the analyzed cases. The second user had a similar background on cardiac anatomy and was provided with the same guidelines on how to perform the corrections. The cases were selected to represent different degree of agreement between \({\text{MBF}}_{{\text{C}}}^{{{\text{vs}}}}\) and \({\text{MBF}}_{{\text{M}}}^{{{\text{vs}}}}\). The DCC obtained after the two manual correction phases (indicated as \({\text{DC}}_{{\text{C}}}^{1}\) and \({\text{DC}}_{{\text{C}}}^{2}\)) were calculated for a total of 12 vessels, three vessel per case.
Statistical analysis
Continuous variables are presented as mean ± SD, while discrete variables are expressed as frequency distributions and percentages. The DC analysis was performed for all pooled vessels and separated by vascular territory and presented as mean values for each group. The presence of statistically significant differences in classification agreements between segmentation modes was tested by way of a Student’s t test, with p < 0.05 as level of significance. The Student’s t test was also used to determine whether the two manual corrections of the anatomical masks provided significantly different classifications with respect to the normal/abnormal ranges.