Deep learning analysis of left ventricular myocardium in CT angiographic intermediate-degree coronary stenosis improves the diagnostic accuracy for identification of functionally significant stenosis

This was a single center, retrospective, observational study. Between 2012 and 2016, 136 consecutive patients who underwent a CCTA within 1 year prior to an invasive FFR measurement were selected. Patients were excluded (n = 10) when automatic LVM segmentation failed or when a significant part was not or incorrectly segmented, following grading defined by Abadi et al [21]. This was the only exclusion criteria used to indirectly exclude scans with bad image quality, suboptimal contrast, motion artifact, or any other artifact. No further in- or exclusion criteria were applied. This study was approved by our local institutional review board at the University Medical Center Utrecht in the Netherlands (protocol number 15/608). The need for informed consent was waived by the institutional review board, since the study was performed on already available imaging data.

All patients underwent CCTA on a 256-slice CT system with a collimation of 128 × 0.625 mm (Brilliance iCT, Philips Healthcare). In accordance with the Society of Cardiovascular Computed Tomography guidelines [22], beta-blockers and/or nitroglycerine was administered to target a heart rate of 60 beats per minute (bpm). Patients were imaged using either an ECG-triggered step-and-shoot protocol (≤ 60 bpm) or a retrospectively ECG-gated spiral protocol (heart rate > 60 bpm). An intravenous contrast agent (70–80 mL) (iopromide, Ultravist 300, Bayer Healthcare), followed by a mix (60–67 mL) of 30% contrast agent and 70% saline, followed by saline (30–40 mL) was administrated using a flow rate of 6 (< 80 kg) or 6.7 mL/s (≥ 80 kg). Bolus tracking with 7-s delay was used for acquisition (threshold of 100 HU for step-and-shoot or 200 HU for spiral). For step-and-shoot, images were acquired at 78% of the R-R interval with a rotation time of 0.27 s. A weight-dependent tube voltage and corresponding current of 80, 100, or 120 kVp and 195, 210, and 210 mAs were used in patient weighing < 50 kg, 50–80 kg, and > 80 kg, respectively. For spiral, a heart rate–dependent pitch (0.16–0.18) and rotation time (0.27–0.33) were used. Weight-dependent tube voltage and corresponding tube current were 100, 120, or 120 kVp and 600, 600, and 700 mAs for patients weighing < 65 kg, 65–80 kg, and > 80 kg, respectively. Images acquired at 75% of the R-R interval were selected for further analysis. For both protocols, axial images were reconstructed with a slice thickness of 0.9 mm at 0.45-mm increment using iterative reconstruction (iDose⁴, Philips Healthcare). Maximum DS for each coronary artery was visually scored in consensus by two observers and categorized in 0%, 1–24%, 25–49%, 50–69%, and ≥ 70% or non-diagnostic, in accordance with dedicated guidelines [23, 24].

All patients underwent invasive coronary angiography (ICA) and FFR according to standard clinical guidelines. FFR measurements were performed in the coronary arteries when deemed clinically indicated. Measurements were performed under adenosine-induced hyperemia (140 μg/kg as continuous intravenous infusion), and FFR was measured as distal as possible in the target vessel using a coronary pressure wire (Pressure™ Certus™ Wire, St. Jude Medical). The FFR value was automatically calculated by dividing the pressure measured distal to the stenosis by the pressure measured at the level of the guiding catheter in mmHg. In case FFR was ≤ 0.80, the stenosis was considered significant. In the vessels with high-grade coronary stenosis at ICA (≥ 90% DS), invasive FFR was not performed, but stenoses were deemed significant based on DS alone [25]. Therefore, when evaluating ischemia on a patient basis, either a significant FFR (≤ 0.8) or high-grade stenosis at ICA (≥ 90% DS) was used to indicate the presence of functionally significant stenoses (Fig. 1).

To evaluate the diagnostic performance of adding LVM DL analysis to evaluation of DS, patients were subdivided into three different groups based on the highest degree of coronary stenosis at CCTA (group 1, DS ≤ 24%; group 2, DS 25–69%, and group 3, ≥ 70%DS). Patients with ≤ 24% DS at CCTA (n = 10) were considered to have a non-significant stenosis, and patients with ≥ 70% DS at CCTA (n = 15) were considered to have a significant stenosis. Patients with intermediate stenosis at CCTA (25–69% DS, n = 101) were referred for fully automatic DL analysis of the LVM to indicate whether a functionally significant stenosis was present (Fig. 2). With the DL method, resting CCTA images were automatically analyzed using the method as described by Zreik et al [13]. The prior technical study [13] focused on development of the DL algorithm, while assessing the effect of the different building blocks. The current clinical study expands and validates the previous study in two ways. First, by employing a clinically more relevant reference standard to indicate functional significant stenosis on a patient level: either a significant FFR (≤ 0.8) or high-grade stenosis at ICA ( ≥90% DS). Second, a combined method of visual stenosis grading on CCTA and only applying the DL-based analysis to the intermediate-degree stenosis is applied. In this way, the DL method could be trained and evaluated on this subset of most challenging patient category of intermediate stenosis only. A full description of the DL method can be found elsewhere [13]; a graphical summary is presented in Fig. 3, and a more detailed description about feature extraction in Supplement Fig. 1. First, LVM was automatically segmented on all CT slices using a multiscale convolutional neural network (Fig. 4). Subsequently, the LVM was characterized (encoded) on all CT slices by the algorithm in an unsupervised manner using a convolutional auto-encoder. Using these encodings, which likely contain information regarding shape, texture, contrast enhancement, and more, features were extracted to represent the whole LVM as a volume. Based on these features, patients were classified according to the presence of functionally significant coronary stenosis using a support vector machine classifier. For the current study, the DL method was trained and validated on the subset of patients with intermediate stenosis (n = 101) using an invasively measured FFR ≤ 0.80 or angiographic high-grade stenosis (≥ 90% DS) as an indicator for presence of functionally significant stenosis. All patients with intermediate stenosis (n = 101) were classified in 10-fold stratified cross-validation experiments. For each 10-fold cross-validation, the dataset was divided into ten stratified subsets whereby one subset was used as validation set and the other nine subsets were used for training. This process was repeated ten times, whereby each subset was used once as validation set. To evaluate robustness of the method, the 10-fold cross-validation was repeated 50 times, with randomization of the data after each repetition. To allow for evaluation of the combined method (DS and LVM DL on intermediate stenosis) on the entire set of patients, patients found to be non-significant (≤ 24% DS) and significant (≥ 70% DS) based on DS on CCTA were assigned probability values of 0 s and 1 s, respectively. For patients with intermediate stenosis, the output probability of the LVM DL model in each cross-validation experiment was employed.

Diagnostic performance of the proposed method (DS evaluation combined with DL analysis of LVM on patients with intermediate stenosis at CCTA) for predicting the presence of functionally significant stenosis on a patient basis was evaluated and compared to the diagnostic performance of DS evaluation alone. Diagnostic performance was evaluated using accuracy, sensitivity, specificity, negative predictive value, positive predictive value, and area under the receiver operating characteristic curve (AUC). For the evaluation of the DS combined with DL analysis, diagnostic performance was evaluated on all 50 cross-validation experiments, and the average ± standard deviation (SD) was used to represent diagnostic performance. For DS only, diagnostic performance is presented with 95% confidence interval (CI). Because prior events and interventions could influence texture changes, an additional sensitivity analysis was performed after exclusion of patients with prior events and interventions (prior myocardial infarction (MI), prior PCI, and/or prior CABG, n = 23). The Shapiro-Wilk test was used to evaluate normality of the data. Continuous values are listed as mean with SD and categorical values as percentages, unless stated otherwise. A p value < 0.05 was used to indicate statistical significance. IBM SPSS version 21.0 (IBM corp.) and MedCalc Statistical Software version 17.7.2 (MedCalc Software BVBA) were used for statistical analyses.

In total, 136 patients were eligible, of which 10 were excluded due to incorrect or failed automatic LVM segmentation. The final study population consisted of 126 patients (77% male, mean age 59 ± 9 years). Baseline characteristics are listed in Table 1.

The median number of days (IQR) between CCTA and invasive FFR was 33 (40). Eighty-one patients (81/126, 64%) had a functionally significant stenosis, of which 69 patients (69/81, 85%) had an FFR ≤ 0.80. In the remaining twelve patients (12/81, 15%), lowest FFR value was > 0.80 and a high-grade stenosis on ICA (≥ 90% DS) was present in another coronary artery (Fig. 1). The distribution of severity of disease on a patient-by-patient basis is depicted in Fig. 5.

Maximum DS on a patient basis was 0%, 1–24%, 25–49%, 50–69%, and ≥70% in 2, 8, 10, 91, and 15 patients, respectively. In 20 patients, at least one segment was scored non-diagnostic, of which 18 patients had a significant stenosis (≥ 50% DS) in another segment. In the remaining two patients, maximum DS in another segment was 25–49%, and PCI was present.

The diagnostic performance of analysis by evaluation of DS alone to indicate functionally significant stenosis was moderate with an AUC of 0.68 (95%CI 0.59–0.76) (Fig. 6). When applying a threshold of ≥ 50% DS to indicate functional significance of a stenosis, sensitivity was 92.6% (95%CI 84.6–97.2%) and specificity 31.1% (95%CI 18.2–46.6%). Changing the threshold to lower or higher DS resulted in commensurate changes in sensitivity or specificity (Table 2). In ten patients (10/126, 8%), the maximum DS at CCTA was ≤ 24% (Fig. 2). In all of these patients, FFR was higher than 0.8 (mean ± SD; 0.89 ± 0.05) and no high-grade stenosis on ICA was present. Fifteen patients (15/126, 12%) had ≥ 70% DS at CCTA, and in 14 of these patients (14/15, 93%), FFR was ≤ 0.8 or high-grade stenosis on ICA was present. In one case, FFR was 0.82, and no high-grade stenosis on ICA was present. The remaining 101 patients with intermediate DS at CCTA were subjected to LVM DL analysis. Classification based on the combination of DS from CCTA (≤ 24% and ≥ 70%) and LVM DL analysis on intermediate stenosis (25–69%) improved discrimination (AUC = 0.76 ± 0.02) compared to classification based on DS from CCTA only (AUC = 0.68 [95%CI 0.59–0.76]) (Fig. 6). The combined method resulted in an increased specificity of 48.4 ± 0.04% at the expense of a small decrease in sensitivity (84.6 ± 0.03%), compared to the specificity 31.1% (95%CI 18.2–46.6%) and sensitivity 92.6% (95%CI 84.6–97.2%) of DS from CCTA only (≥ 50% DS to indicate functional significant stenosis) (Table 2). To evaluate the robustness of the DL method applied on different patient cohorts, the DL analysis was applied on the complete cohort (n = 126) and compared to DL analysis of intermediate stenosis only (n = 101) (supplement Fig. 2).

In eight patients (8/103, 8%), maximum DS at CCTA was ≤ 24%, and in 12 patients (12/103, 12%), maximum DS was ≥ 70%. DL was applied to the intermediate stenosis (n = 83). Also in this sub analysis, the combined method showed an improved discrimination (AUC = 0.77 ± 0.02) compared to evaluation of DS alone (AUC = 0.67 [95%CI 0.57–0.76]) (Supplement Fig. 3). Other measures of diagnostic performance were also comparable to the results of the full sample (Supplement Table 1).