Evaluation of transfer learning in deep convolutional neural network models for cardiac short axis slice classification

In computer-aided analysis of cardiac MRI data, segmentations of the left ventricle (LV) and myocardium are performed to quantify LV ejection fraction and LV mass, and they are performed after the identification of a short axis slice coverage, where automatic classification of the slice range of interest is preferable. Standard cardiac image post-processing guidelines indicate the importance of the correct identification of a short axis slice range for accurate quantification. We investigated the feasibility of applying transfer learning of deep convolutional neural networks (CNNs) as a means to automatically classify the short axis slice range, as transfer learning is well suited to medical image data where labeled data is scarce and expensive to obtain. The short axis slice images were classified into out-of-apical, apical-to-basal, and out-of-basal, on the basis of short axis slice location in the LV. We developed a custom user interface to conveniently label image slices into one of the three categories for the generation of training data and evaluated the performance of transfer learning in nine popular deep CNNs. Evaluation with unseen test data indicated that among the CNNs the fine-tuned VGG16 produced the highest values in all evaluation categories considered and appeared to be the most appropriate choice for the cardiac slice range classification.

Prognostic Value of Non-Ischemic Ring-Like Left Ventricular Scar in Patients with Apparently Idiopathic Non-Sustained Ventricular Arrhythmias

Background: Left ventricular (LV) scar on late gadolinium enhancement (LGE) cardiac magnetic resonance (CMR) has been correlated with life-threatening arrhythmic events in patients with apparently idiopathic ventricular arrhythmias (VA). We investigated the prognostic significance of a specific LV-LGE phenotype characterized by a ring-like pattern of fibrosis. Methods: 686 patients with apparently idiopathic non-sustained VA underwent contrast enhanced CMR. A ring-like pattern of LV scar was defined as LV subepicardial/midmyocardial LGE involving at least 3 contiguous segments in the same short-axis slice. The endpoint of the study was time to the composite outcome of all cause death, resuscitated cardiac arrest due to ventricular fibrillation (VF) or hemodynamically unstable ventricular tachycardia (VT) and appropriate implantable cardioverter defibrillator therapy. Results: A total of 28 (4%) patients had a ring-like pattern of scar (Group A), 78 (11%) a non ring-like pattern (Group B), and 580 (85%) had normal CMR with no LGE (Group C). Group A patients were younger compared to Group B and Group C (median age 40 vs. 52 vs. 45 years, p<0.01), more frequently males (96% vs. 82% vs. 55%, p<0.01) with a higher prevalence of family history of sudden cardiac death/cardiomyopathy (39% vs. 14% vs. 6%; p<0.01), and more frequent history of unexplained syncope (18% vs. 9% vs. 3%, p<0.01). All patients in Group A showed VA with a right bundle branch block morphology vs. 69% in Group B and 21% in Group C (p<0.01). Multifocal VA were observed in 46% of Group A patients compared to 26% of Group B and 4% of Group C (p<0.01). After a median follow-up of 61 (34-84) months, the composite outcome occurred in 14 patients (50%) in Group A vs. 15 (19%) in Group B and 2 (0.3%) in Group C (p<0.01). After multivariable adjustment, the presence of LGE with ring-like pattern remained independently associated with increased risk of the composite endpoint (HR 68.98, 95% CI 14.67-324.39, p<0.01). Conclusions: In patients with apparently idiopathic non-sustained VA, nonischemic LV scar with a ring-like pattern is associated with malignant arrhythmic events.

Comparison of the T2 effects of three T1 mapping sequences of the myocardium at 3T and 1.5T

Funding Acknowledgements Type of funding sources: None. Purpose To compare the T2 effects of three T1 mapping sequences of the myocardium at 3T and 1.5T. Materials and Methods We measured native T1 values by three T1 mapping sequences (FFM, MPI and LONG) and T2 values of 630 consecutive subjects (patients and healthy volunteers) in a mid-ventricular short axis slice by regions of interest (ROIs) placed conservatively within the septal myocardium. Correlations between myocardial T1 and T2 values were analyzed. Results Native T1 values differed significantly depending on the sequence, with MPI providing consistently higher mean values than FFM and LONG, LONG providing higher mean values than FFM (all p &lt; 0.001). T1 by FFM, MPI and LONG in the total population, disease group, non-ischaemic cardiomyopathy group were all weakly related to T2 at 3T. The correlation coefficient of MPI were the highest in the total population and disease group, but there is no significant difference in correlation coefficients (all p &gt; 0.05). Similarly, T1 by FFM, LONG in the total population, disease group, NICM group were all weakly-moderately related to T2 at 1.5T. In the control group, only T1 by MPI was moderately and positively related to T2(r = 0.469, p = 0.037). In the ischaemic cardiomyopathy group, only T1 by MPI was weakly and positively related to T2 (r = 0.334, p = 0.011). In the stress condition, T1 by FFM and MPI was weakly and positively related to T2 (r = 0.280,0.210, p = 0.001,0.012, respectively). Conclusion The T2 effects existed in three T1 mapping sequences of the myocardium in varying degrees, we should consider the potential bias from T2 when analyzing the abnormal T1 values of myocardium. Abstract Figure.

Myocardial Mapping in Systemic Sarcoidosis: A Comparison of Two Measurement Approaches

Purpose To investigate if T1 and T2 mapping is able to differentiate between diseased and healthy myocardium in patients with systemic sarcoidosis, and to compare the standard mapping measurement (measurement within the whole myocardium of the midventricular short axis slice, SAX) to a more standardized method measuring relaxation times within the midventricular septum (ConSept). Materials and Methods 24 patients with biopsy-proven extracardiac sarcoidosis and 17 healthy control subjects were prospectively enrolled in this study and underwent CMR imaging at 1.5 T including native T1 and T2 mapping. Patients were divided into patients with (LGE+) and without (LGE–) cardiac sarcoidosis. T1 and T2 relaxation times were compared between patients and controls. Furthermore, the SAX and the ConSept approach were compared regarding differentiation between healthy and diseased myocardium. Results T1 and T2 relaxation times were significantly longer in all patients compared with controls using both the SAX and the ConSept approach (p < 0.05). However, LGE+ and LGE– patients showed no significant differences in T1 and T2 relaxation times regardless of the measurement approach used (ConSept/SAX) (p > 0.05). Direct comparison of ConSept and SAX T1 mapping showed high conformity in the discrimination between healthy and diseased myocardium (Kappa = 0.844). Conclusion T1 and T2 mapping may not only enable noninvasive recognition of cardiac involvement in patients with systemic sarcoidosis but may also serve as a marker for early cardiac involvement of the disease allowing for timely treatment. ConSept T1 mapping represents an equivalent method for tissue characterization in this population compared to the SAX approach. Further studies including follow-up examinations are necessary to confirm these preliminary results. Key Points: Citation Format

P3096The cardioprotective effect of FFR-significant multivessel disease detected by cardiac magnetic resonance imaging in patients following ST-segment elevation myocardial infarction. Results from DANAMI3

Background In patients with ST-segment elevation myocardial infarction (STEMI) treated with primary percutaneous coronary intervention (PCI), reperfusion injury accounts for a significant part of the final infarct size, which is directly related to patient prognosis. In animal studies brief periods of ischemia in non-infarct related coronary arteries protects the myocardium via remote ischemic perconditioning. Fractional flow reserve (FFR) measures functional significant coronary stenosis which may offer remote ischemic perconditioning of the myocardium. It has not previously been investigated if FFR-significant stenosis in non-culprit myocardium offers cardioprotection following STEMI. Purpose To investigate cardioprotective effect of FFR-significant multivessel disease (MVD) on final infarct size and myocardial salvage in a large contemporary cohort of patients with ST-segment elevation myocardial infarction (STEMI). Methods and results We included 509 patients with STEMI from the DANAMI-3 trial, divided into three groups: 388 (76%) patients had single vessel disease (SVD), 34 (7%) had non-FFR-significant MVD and 192 (17%) had FFR-significant MVD. CMR was performed at baseline and three months after primary PCI. There was no difference in final infarct size; mean infarct size (% left ventricular mass) SVD 9±3%; non-FFR-significant MVD 9±3%; and FFR-significant MVD 9±3%, p=0.95, or in myocardial salvage index (MSI) between groups, calculated as (area-at-risk – infarct size)/area-at-risk; mean index (%) SVD 67±23%; non-FFR-significant MVD 68±19%; and FFR-significant MVD 67±21%, p=0,99. In multivariable regression analyses FFR-significant MVD was not associated med larger MSI (p=0.84) or lower infarct size (p=0.60). Figure 1. A. Late gadolinium (LGE) cardiac magnetic resonance (CMR) image of a mid-ventricular short-axis slice. Hyperintense signals (arrow) shows contrast enhancement in the anterior-septal segments, indicating myocardial infarction (MI). B. Same patient. T2-weighted image of the same mid-ventricular short-axis slice. Hyperintense signals (arrows) shows edema in the anterior-septal segments. Conclusions FFR-significant functional MVD of non-culprit myocardium does not offer cardioprotection in patients following STEMI.

Dilated cardiomyopathy and arrhythmogenic left ventricular cardiomyopathy: a comprehensive genotype-imaging phenotype study

Aims Myocardial scar detected by cardiovascular magnetic resonance has been associated with sudden cardiac death in dilated cardiomyopathy (DCM). Certain genetic causes of DCM may cause a malignant arrhythmogenic phenotype. The concepts of arrhythmogenic left ventricular (LV) cardiomyopathy (ALVC) and arrhythmogenic DCM are currently ill-defined. We hypothesized that a distinctive imaging phenotype defines ALVC. Methods and results Eighty-nine patients with DCM-associated mutations [desmoplakin (DSP) n = 25, filamin C (FLNC) n = 7, titin n = 30, lamin A/C n = 12, bcl2-associated athanogene 3 n = 3, RNA binding motif protein 20 n = 3, cardiac sodium channel NAv1.5 n = 2, and sarcomeric genes n = 7] were comprehensively phenotyped. Clustering analysis resulted in two groups: ‘DSP/FLNC genotypes’ and ‘non-DSP/FLNC’. There were no significant differences in age, sex, symptoms, baseline electrocardiography, arrhythmia burden, or ventricular volumes between the two groups. Subepicardial LV late gadolinium enhancement with ring-like pattern (at least three contiguous segments in the same short-axis slice) was observed in 78.1% of DSP/FLNC genotypes but was absent in the other DCM genotypes (P < 0.001). Left ventricular ejection fraction (LVEF) and global longitudinal strain were lower in other DCM genotypes (P = 0.053 and P = 0.015, respectively), but LV regional wall motion abnormalities were more common in DSP/FLNC genotypes (P < 0.001). DSP/FLNC patients with non-sustained ventricular tachycardia (NSVT) had more LV scar (P = 0.010), whereas other DCM genotypes patients with NSVT had lower LVEF (P = 0.001) than patients without NSVT. Conclusion DSP/FLNC genotypes cause more regionality in LV impairment. The most defining characteristic is a subepicardial ring-like scar pattern in DSP/FLNC, which should be considered in future diagnostic criteria for ALVC.

Role of T1 Mapping As a Complementary Tool to T2* for Cardiac Iron Overload Assessment

Introduction. Iron overload-related heart failure is the principal cause of death in transfused Thalassemia Major (TM) (Modell B, Cardiovasc Magn Reson 2008;10:42-48). Iron toxicity is dose dependent so a strategy of chelation therapy titration (Kirk P, Circ 2009;120:1961-1968) before the onset of left ventricle (LV) impairment changes outcomes (AlpenduradaF, Eur Heart J. 2010; 31:1648-54). The presence of iron in tissue detectably changes the magnetic properties of water, T1, T2 and T2*, as validated against tissue in animal and human models (Carpenter JP, Circ 2011;14:1519-28). T2*, the most used technique, is susceptible to non-iron influence (susceptibility artefact) and has low accuracy for high and low iron levels (Carpenter JP, J Cardiovasc Magn Reson. 2014;12:16-62). T1 mapping could complement T2* as it appears to have superior reproducibility and to detect mild iron missed by T2* (Abdel-Gadir A. J Cardiovasc Magn Reson. 2015;17(Suppl1):P312. Sado DM, J Magn Reson Imaging. 2015;41:1505-11), but studies to date have been small and not using state-of-the-art sequences. Methods. In a prospectively single centre study of 138 TM patients and 32 healthy volunteers (HV) (no known medical conditions, normal CMR scan), we compared T1 mapping (Modifier Lock Locker Inversion sequence - MOLLI - Siemens Works in progress 448B) to the gold-standard dark (DB) and bright (BB) blood T2*, acquired on an Avanto 1.5T (Siemens Healthcare, Erlangen, Germany). For both T2* sequences, a single 10mm mid-ventricular short axis slice was imaged at 8 echo-times (2.58ms to 18.19ms, increment 2.23ms), flip angle=20¡, FOV read/phase=400mm/56,3%. The same slice was used for T1 images (thickness 6mm, distance factor=67%, FOV read/phase=360/75%, TR=740, TE=1.13, with motion correction for the in-line map generation). Results and discussion.All participants provided informed consent. Table1 illustrates patients' and HV's details. T2* was defined normal under the cutpoint value of 20ms. T1 normal range, defined by the HV cohort was 918-1015ms (the 2.5-97.5 quantiles with CI 95%). For DBT2*<20ms, both BBT2* and T1 mapping were broadly indistinguishable from DBT2* (DBT2* vs BBT2* R2=0.95; DBT2* vs T1 R2=0.92; all p<0.001). All subjects with low DBT2* (n=24, 17.4%) had low T1; 52 patients had normal DBT2* but low T1 mapping, i.e. 38% patients were reclassified from normal to iron loaded by T1. The relationship between DBT2* and MOLLI was described by a log-log linear regression (R2=0.80, p<0.001). Upper panel of Fig1 shows T1 vs DBT2* correlation over a 20ms window as the window moves by 1ms at a time on X-axis (so at X-axis point 'n', the Y-value is the R2 of the correlation of DBT2* vs T1 over the range n-to-n+20ms). As shown by lower panel of Fig1, three domains can be observed: strong relationship in the T2*=0-20ms range (R2=0.92, p<0.001); good relationship in the 21-28ms range, where the curve depicts a plateau (R2=0.80-0.77, p<0.001) and no relationship above 28ms. Given the conservative approach used to set T2* normality as above 20ms (Carpenter JP, Circ 2011;14:1519-28), the evidence that T2* SD values increase even for borderline T2* mean values (~20ms) (Anderson LJ, Eur Heart J. 2001;22:2171-9), and that 39% of normal T2* subjects have a low T1, we support prior suggestions that T1 is detecting mild iron in most of the subjects with DBT2*=20-28ms, missed by T2* as the threshold has had to be set too low for sensitivity reasons (Sado DM, J Magn Reson Imaging. 2015;41:1505-11). T1 mapping is thus a useful complementary tool to T2* both for clinical and research purposes. The reported reproducibilities of T1 square for power calculations and would translate into 6.25 to 49 times more power in studies to detect iron change (Alam MH, J Cardiovasc Magn Reson. 2015;24:17-102). Colour maps make iron instantly visible and add a confirmation step. Whether mild iron missed by T2* is important is unclear. In our cohort, 24-months follow-up was available for 9 patients with normal DBT2* and low T1. Although no statistical consideration is possible due to the small number, in those patients an increase in LV end diastolic volume was observed (from 78±18ml to 84±15ml), suggesting possible cardiotoxicity of even mild amount of iron. Further work is needed, especially in frail cohorts like children starting chelation and around pregnancy. Table 1 Table 1. Figure 1 Figure 1. Disclosures Moon: gsk: Consultancy; Genzyme: Research Funding; Shire: Membership on an entity's Board of Directors or advisory committees. Cappellini:Novartis: Membership on an entity's Board of Directors or advisory committees; Genzyme-Sanofi: Membership on an entity's Board of Directors or advisory committees; Celgene: Membership on an entity's Board of Directors or advisory committees.

Abstract 19100: Transmural and Lateral Remodeling of the Post-infarction Scar Tissue: Serial Cardiac Magnetic Resonance Imaging Study From the Metocard-Cnic Trial

Background: The remodeling process that occurs following an acute myocardial infarction produces changes in the myocardial scar and the surrounding tissue. The spatial evolution of the scar has not been yet characterized using MRI. Purpose: To describe the spatial behavior of myocardial scar on its transmural and lateral dimensions following an acute myocardial infarction. Methods: A total of 220 patients with acutely reperfused anterior STEMI (METOCARD-CNIC trial population) were studied. All the patients underwent cardiac MRI at day 7 and 6 months after presentation. The spatial distribution of the scar was analyzed using short axis late gadolinium enhancement images. Endo and epicardial contours of each LV short axis slice was traced, and each one was divided by 100 cords for analysis (Figure 1A). The lateral extension was defined as the percentage of cords with enhancement in every slice, and the transmural extension as the percentage of enhancement within each cord, nested in the lateral extension (Figure 1B). All slices were weighed according to their relative mass. Data were compared by paired t test. Results: Six months after STEMI, myocardial scar was smaller in both dimensions (Figure 1C). Mean ± SEM lateral scar, as a percentage of cords affected, at 7 days and 6 months were, respectively, 27.9 ± 1.1 and 22.2 ± 1.0 (p < 0.001). Transmurality also decreased significantly, mean percentage at day 7 was 46.8 ± 1.5 and at 6 months 35.4 ± 1.4 (p < 0.001). All these changes were accompanied by an increase in LV end-diastolic volume (171.47 ± 2.5 and 190.61 ± 2.9 (p < 0.001)). Conclusions: The myocardial scar remodeling process following a STEMI includes a reduction in both its transmural and lateral extensions.

Abstract 18987: Wavefront Progression of Necrosis During Myocardial Infarction Revisited: The Wave Expands From Endo to Epi as Well as From Center to Lateral Borders

Background and purpose: The widely accepted wavefront theory of myocardial necrosis during acute coronary occlusion describes its progression as a wavefront from endo towards epicardium. The lateral affection of the myocardium would be predetermined by the location of the coronary occlusion. At the time this phenomenon was firstly described, advanced imaging technology was not available. Current cardiac magnetic resonance (CMR) is able to accurately characterize the necrosis progression in vivo. Methods: A total of 220 patients with reperfused anterior STEMI (METOCARD-CNIC trial population) were studied. All the patients underwent cine, T2-weighted and late gadolinium enhancement CMR at day 5-7 after STEMI. Endo- and epicardial contours of each LV short axis slice was traced, and each one was divided by 100 cords for analysis. Infarct size was characterized in its 2 dimensions: lateral (defined as the percentage of cords with enhancement in every slice) and transmural extension (percentage of enhancement within each cord, nested in the lateral extension). All slices were weighed according to their relative mass. We compared these parameters between the cardioprotected metoprolol group and the control group. Results: A strong lineal correlation between transmural and lateral infarct extension was observed (r: 0.88, p<0.001). We found that despite the edema lateral extension was larger than the IS lateral extension, this difference decreased as the transmurality affection increased. Finally, the infarct reduction exerted by metoprolol was derived from both the transmural and lateral extensions of IS. Conclusions: The transmural and lateral extension of necrosis after myocardial infarction are directly correlated. A proven cardioprotective therapy prevented the necrosis progression in both dimensions. These findings cannot be explained by the classical wavefront theory, suggesting that the necrotic wavefront progresses also in the lateral direction.