Semi-quantitative Analysis of Adenosine Perfusion Magnetic Resonance Imaging Outperforms Visual Analysis in Identifying Fractional Flow Reserve-Altered Intermediate-grade Coronary Artery Stenosis: A Pilot Study

Purpose To evaluate the diagnostic accuracy of semi-quantitative adenosine perfusion magnetic resonance imaging(MRI) to determine fractional ow reserve(FFR) ≤ 0.80 intermediate-grade coronary stenoses as compared to visual analysis. Methods Forty-six patients (mean age 61±9 years;33 males) with 49 intermediate-grade stenoses underwent adenosine perfusion MRI and FFR measurement within 4 months between 2010 and 2013. Retrospective interpretation of all prospectively acquired data was performed. MRI was visually assessed by 2 experienced readers twice with one-year interval, the second time with the knowledge of the diseased artery. All myocardial enhancement maximal upslopes were evaluated distal to the coronary stenosis (=RISK) and in remote myocardium supplied by normal arteries (=REMOTE); stress subendocardial relative myocardial perfusion index (RMPI; RISK/REMOTE upslopes) was assessed in predicting FFR ≤ 0.80 stenoses. Deep learning boosting models including all RISK and REMOTE upslopes were tted to conrm the added value of accounting for perfusion changes in remote myocardium for FFR prediction. and (31%) Both moderate accuracy (range: 36/49(73%)-38/49(78%)) ≤ even a cutoff (43/49(88%)) individual ≤ for REMOTE perfusion of 44/49(90%) FFR ≤ 0.80.


Introduction
Coronary computed tomography angiography and catheter coronary angiography poorly predict ow limitation, especially for stenoses in the intermediate-grade range (i.e. 40%-70% diameter reduction) [1,2] that may represent up to 42% of coronary stenoses [3]. Additional functional assessment is often required to guide therapeutic management as approximately only one-third of patients with intermediate-grade stenoses suffer from ischemia and would bene t from revascularization [4,5].
Invasive fractional ow reserve (FFR) measurement is the standard of reference for the functional signi cance (ischemia) of coronary stenoses. However, its use as a rst step in intermediate-grade lesions is prevented by its invasiveness, the use of ionizing radiation, time and the costs of pressure wires [6,7]. Moreover, the use of invasive FFR varies widely depending on the practice of interventional cardiologists.
While single-photon emission computed tomography [8] and dobutamine stress echocardiography [9] have moderate accuracy (66%-72%) for identifying FFR-altered (i.e. FFR≤0.8) intermediate-grade stenoses, no study has speci cally addressed this subgroup of stenoses using stress perfusion magnetic resonance imaging (MRI). Though it has a higher accuracy for the detection of myocardial ischemia as compared to other non-invasive imaging modalities [10], visual analysis of adenosine perfusion MRI in daily clinical practice may be misleading compared to FFR. Actually, MRI assesses perfusion defects, interrogating both the epicardial coronary artery stenosis and the downstream microvasculature, whereas the FFR value is inherently corrected for the microvascular resistance [11]. The relative myocardial perfusion index (RMPI), recently-described on MRI as the ratio of the maximal enhancement upslope distal to a coronary artery stenosis to that of a normally perfused area during stress perfusion is similar to the FFR approach [12].
As this semi-quantitative index better correlated with the FFR value than the uncorrected enhancement upslope [12], we hypothesized that RMPI could provide high diagnostic accuracy for the detection of FFR-altered intermediate coronary stenoses. Accordingly, the aim of this study was to evaluate the diagnostic accuracy of stress RMPI to determine FFR≤0.80 intermediate-grade coronary stenoses as compared to individual visual analysis.

Patients and study protocol
This study protocol was approved by the local institutional ethics committee, and patients provided written informed consent. Between 2010 and 2013, consecutive patients with an intermediate-grade stenosis on computed tomography angiography involving one or two major epicardial coronary vessels >1.5mm in diameter were eligible for a study requiring both catheter coronary angiography with FFR measurements and adenosine perfusion MRI within 4 months as previously reported [12]. MRI examinations were performed on a 1.5T MR scanner (Avanto, Siemens Healthineers), as previously reported [12]. In short, the examination consisted in performing stress and resting perfusion on dynamic contrast-enhancement imaging (each using 0.1 mmol/kg Gadodiamide, Omniscan Ò ), and late-gadolinium enhancement (LGE) in the same three short-axis positions (see the supplementary materials for more details). These prospectively acquired data were analyzed retrospectively as follows: visually twice with one-year interval, the second reading with knowledge of the diseased coronary artery; semi-quantitatively and using machine learning analysis to con rm the added value of accounting for perfusion changes in remote myocardium in predicting the FFR value.

MRI analysis Visual analysis
Two readers (AN, JND) with more than 10-years of experience in cardiac MRI, blinded to patient's characteristics, history and coronary angiography and FFR ndings, performed twice an individual visual analysis of perfusion MRI, using dedicated software (Syngo Via Ò , Siemens Health). First, splenic switch-off was qualitatively assessed to evaluate the appropriateness of the vasodilatation response after adenosine administration [13] . Myocardial ischemia was de ned as stress-induced myocardial perfusion defect in the absence of LGE in the same segment, as previously reported [3]. The readers had no common training before the study and had freedom to adjust the display window level and width. In a nal step, all reading discordances were solved by consensus. Twelve months after the rst readings, a second round of individual and consensus visual analysis was performed by the same readers who then were provided with full knowledge of the coronary stenosis location, but still blinded to the FFR data.

Semi-quantitative analysis
Semi-quantitative analysis was performed under the supervision of an expert in cardiac MRI using dedicated software (MOCO, Syngo ViaVA30 Ò , cardiac Engine-perfusion module, Siemens Healthineers). This operator had full knowledge of the location of the stenosis, but was blinded to the FFR data. He performed a visual analysis using the same scheme as the 2 other readers to determine areas of myocardial perfusion defects.
As previously described [12], equally divided subendocardial (END) and subepicardial (EPI) regions of interest and timesignal intensity curves were obtained during adenosine stress in the myocardium distal to the stenosis (=RISK). When the RISK area involved more than one segment of the left ventricle representation, the myocardial segment with the greatest lateral and transmural extent of the perfusion defect was used for further measurements. Then, similar curves were obtained for a remote myocardial segment without a stenosis ≥40% diameter reduction on the supplying artery on QCA (=REMOTE) (Figure 1).
When no myocardial perfusion defect was visualized, the RISK segment was de ned distal to the anatomic location of the coronary stenosis and the remaining steps were performed as when a perfusion defect could be visually detected. In patients with more than one intermediate-grade stenosis, each corresponding area of myocardial supply was assessed separately. Subsequently, these regions of interest were copy-pasted on the resting perfusion images. If necessary, manual correction was made to adjust the region of interest placement.
The stress subendocardial RMPI (i.e.: RISK/REMOTE mean maximal enhancement upslopes) of each stenosis was assessed for the diagnosis of FFR≤0.80 stenosis, as previously reported [12].
Maximal enhancement upslopes derived from subendocardial and subepicardial time-signal intensity curves per RISK and REMOTE areas (n=8, Figure 1) for each of the stenoses were normalized to the respective left ventricle cavity enhancement maximal upslope and used for Boosting machine learning.

Statistical analysis
Statistical analyses were performed using R (version 3.2.3, with the model-based boosting package 2.6-0). Normally distributed continuous variables are expressed as mean +/-standard deviation (SD). Comparisons between continuous variables were performed using two-tailed Student t-tests, and comparisons of proportions were performed using χ² tests.
A regression model was tted to determine the best cut-off value for stress subendocardial RMPI in predicting FFR≤0.80. Then, boosting models including all semi-quantitative parameters collected in the RISK and REMOTE myocardium were used to predict FFR-related outcomes and to validate the added value of accounting for perfusion parameters in REMOTE myocardium in the prediction of these outcomes. As more extensively described in the supplementary materials, the Boosting models were assessed with the Akaike Information Criterion (AIC) value.
Subsequently, the best model was optimized for a categorical outcome (i.e.: predicting FFR≤0.80). The diagnostic values were expressed as sensitivity, speci city, positive predictive value, negative predictive value, likelihood ratios and accuracy. The diagnostic accuracies were compared between visual readings, stress subendocardial RMPI and the boosting predictive model for FFR≤0.80, using binomial exact tests. P-values<0.05 were considered to express a statistically signi cant difference.

Bringing in information from the perfusion in remote myocardial areas on top of information in the RISK areas beyond
the stenosis was suggested to improve the predictive performance (AIC from -100 to -140). The boosting model computing all semi-quantitative perfusion MRI parameters con rmed their importance with a signi cantly higher diagnostic accuracy (44/49, 90%) compared to all individual visual readings in predicting FFR ≤ 0.80 stenoses, regardless of the prior knowledge of the stenosis localization (all p-values <0.05). Furthermore, this model tends to ful ll the criteria for a good diagnostic test, with positive and negative likelihood ratios of 9.82 and 0.15, respectively (Table 3). The proportions by which the percentages were calculated are given in parentheses.  Single-observer visual analysis identi ed FFR≤0.80 intermediate-grade stenosis with a moderate accuracy and highly variable sensitivities and speci cities, likely owing to the freedom in image setting adjustment and dark-rim artifact assessment. The agreement would have been higher, but less representative of the "real life" if a pre-study training of the readers would have been organized [14]. These potential reading pitfalls suggest that the interpretation of myocardial signal abnormality depends on many more factors beyond the reader's experience.
Consensus reading resulted in higher diagnostic accuracy for FFR≤0.80 stenosis while keeping sensitivity and speci city close to the individual reader's highest levels. These results are lower than previous perfusion MRI studies [3,15,16], which can be explained by the exclusive inclusion of intermediate-grade stenoses in our study. It is also in line with the reported lower sensitivity of MRI in identifying FFR≤0.80 lesions in a subanalysis of intermediate-grade stenoses from a larger series [17]. Only few studies have speci cally addressed coronary ow-limitation in intermediate-grade stenoses using other non-invasive techniques such as stress dobutamine MRI [18], dobutamine stress echocardiography [9], and single-photon emission tomography [8,19,20], with respective sensitivity and speci city ranges of 62%-95% and 69%-90%, all con rming the challenge posed by this range of stenoses. Even the knowledge of the area-at-risk did not increase the accuracy of visual readings. Indeed, the perfusion defects induced by intermediate-grade stenoses are likely to be shallower and less extended, thus more di cult to perceive and to distinguish from subendocardial dark-rim artefacts, than those caused by high-grade stenoses [21,22]. This implies that beyond encouraging consensus reading of perfusion MRI to mitigate the reader's perception biases, diagnostic use of MRI as a gatekeeper to predict functional signi cance of intermediate-grade stenoses demands improvement.
Actually, visual analysis assesses only perfusion defects beyond a coronary stenosis, and does not account for perfusion in normal perfusion areas, in contrast to the FFR value [23]. As expected, the accuracy of the FFR prediction using deep-learning statistics was higher when accounting for perfusion parameters in remote, normal myocardium than when these parameters were not included. These predictive models were used as proof of concept of the added value of accounting for perfusion parameters in remote myocardium for FFR prediction. It supports the need of an integrative interpretation of the entire myocardium for improved perfusion MRI assessment in determining FFR≤0.80 stenoses, but is di cult to implement in clinical practice.
In this setting, the reported stress subendocardial RMPI [12] provides a simpli ed and more useful semi-quantitative parameter for clinical practice, with a high diagnostic accuracy to determine FFR≤0.80 intermediate-grade stenoses, in line with those of previous meta-analyses, including mainly semi-quantitative and quantitative MRI analyses [10]. This approach has also been reported in stress dynamic computed tomography perfusion showing better accuracy to identify ow-limiting stenoses than the myocardial blood ow in the area-at-risk [24]. Nevertheless, using RMPI three false-negative and three false-positive cases remained for the FFR≤0.8 cutoff, owing to image artifacts and the existence of the so-called gray-zone of FFR values (0.75-0.80) [25]. In addition, Ghekiere et al. reported in their series that all three false-negatives for RMPI exhibited splenic switch-off on adenosine imaging (i.e. did not receive appropriate adenosine vasodilatation) [12].
Our study has certain limitations including the relatively low number of patients and FFR≤0.80 stenosis. Second, the MRI studies were performed between 2010 and 2013 with an older generation equipment. Both visual and deep learning results could have been improved using a higher resolution adenosine perfusion MRI as it improves the detection of subendocardial ischemia [26]. Third, substantial amount of dropouts occurred after QCA to maintain stenoses within the intermediate-grade range and to control the possible bias related to the hemodynamic interactions between distinct coronary territories and that of successive stenosis. We nevertheless included a su cient number of patients regarding the sample estimation for statistical signi cance, and approximately one-third of the intermediategrade stenoses had an FFR≤0.80, as reported in the literature. Larger cohorts of patients will be suitable to con rm and validate the results of our study. Finally, in spite of epicardial stenosis, several confounders may alter myocardial perfusion on MRI [27] and, therefore, its diagnostic values in predicting ow-limitation as de ned by the FFR value. Systematic bias such as cardiac-phase variability of the myocardial perfusion was not taken into account [28,29], but can be ignored as long as a single-slice frame is evaluated, as it was done in this study. Patient-related confounders inherently limit the validity of our data to a population of individuals with similar cardiovascular risk factors for microvascular disease.
In conclusion, consensus reading should be encouraged in clinical practice to improve the diagnostic accuracy of visual analysis of adenosine perfusion MRI in predicting FFR≤0.80 intermediate-grade coronary stenoses. Semiquantitative analysis using RMPI has a higher diagnostic accuracy than individual visual analysis, but further studies with larger patient samples are needed to con rm its clinical value as a gatekeeper for invasive FFR in patients with intermediate-grade stenoses. Declarations