Deep Learning for Diagnosis of Chronic Myocardial Infarction on Nonenhanced Cardiac Cine MRI

D of chronic myocardial infarction (MI) is an important clinical task because the management of and treatment planning for patients is different for chronic MI versus acute MI (1,2). The extent of chronic MI, including location, size, and transmurality, provides rich information for patient diagnosis, prognosis, and therapy planning (3). Therefore, accurate delineation and comprehensive evaluation of chronic MI is of great clinical interest. Late gadolinium enhancement (LGE) MRI has been established as the ground truth reference technique for chronic MI evaluation (4–6). However, including LGE MRI in the MRI examination extends the scanning duration and there are also growing concerns about its safety (7–9). While LGE MRI is contraindicated in patients with severe renal impairment, a recent study has also shown that gadolinium might deposit into the skin, dentate nucleus, and globus pallidus of patients with normal renal function (10). A reliable technique to detect and delineate MI without the need for gadolinium-based contrast agent would therefore be highly desirable. T1 and T2 mapping techniques (11) are non–contrast material–enhanced approaches that show longer T1 and T2 relaxation times in acute MI compared with normal myocardium. In comparison, while T1 relaxation time is Deep Learning for Diagnosis of Chronic Myocardial Infarction on Nonenhanced Cardiac Cine MRI

D iagnosis of chronic myocardial infarction (MI) is an important clinical task because the management of and treatment planning for patients is different for chronic MI versus acute MI (1,2). The extent of chronic MI, including location, size, and transmurality, provides rich information for patient diagnosis, prognosis, and therapy planning (3). Therefore, accurate delineation and comprehensive evaluation of chronic MI is of great clinical interest.
Late gadolinium enhancement (LGE) MRI has been established as the ground truth reference technique for chronic MI evaluation (4-6). However, including LGE MRI in the MRI examination extends the scanning duration and there are also growing concerns about its safety (7)(8)(9). While LGE MRI is contraindicated in patients with severe renal impairment, a recent study has also shown that gadolinium might deposit into the skin, dentate nucleus, and globus pallidus of patients with normal renal function (10). A reliable technique to detect and delineate MI without the need for gadolinium-based contrast agent would therefore be highly desirable.
T1 and T2 mapping techniques (11) are non-contrast material-enhanced approaches that show longer T1 and T2 relaxation times in acute MI compared with normal myocardium. In comparison, while T1 relaxation time is longer in chronic MI than in normal myocardium, resolution of edema within the infarct results in no significant difference in T2 (8,12,13). However, while reproducible (8,14), measured relaxation times are protocol and field-strength specific and normal ranges are broad (15)(16)(17). These techniques also require the acquisition of additional multiple breath-hold data sets with appropriate MRI sequences and extend overall examination duration. Alternatively, MRI feature tracking is an approach that differentiates regional myocardial morphology and cardiac wall motion abnormalities resulting from MI (18,19) on non-contrast-enhanced cardiac cine MRI, which is acquired as part of a standard cardiac MRI examination. However, only the presence and position information of the MI can be extracted from these examinations and these techniques may be further limited by the need for time-consuming processing after the examination.
We propose a fully automatic framework for chronic MI delineation via deep learning on nonenhanced cardiac cine MRI and assess its accuracy for identifying the presence, position, transmurality, and size of the MI without the need for gadolinium injection.

Materials and Methods
This retrospective study was approved by our institutional review board in accordance with local ethics procedures. Further consent was waived with approval.

Patients
Detailed demographics and left ventricle volumetric data are summarized in Table 1. Between October 2015 and March 2017, 212 patients with chronic MI (based on clinical symptoms, electrocardiogram changes, and greater than twofold elevation of creatine kinase and/or positive troponin T) and 87 control patients (without negative LGE MRI) were selected from a single center for retrospective inclusion in our study (Fig 1, Appendix E1 [online]). Major exclusion criteria were acute MI, angina without MI, all kinds of nonischemic cardiomyopathy, cardiac neoplasm, valvular heart disease, congenital Abbreviations AUC = area under the receiver operating characteristic curve, CI = confidence interval, LGE = late gadolinium enhancement, MI = myocardial infarction Summary Deep learning on nonenhanced cardiac MRI data can detect the presence and extent of chronic myocardial infarction. This approach may have potential to reduce use of gadolinium contrast administration.

Key Points
n A deep learning method to identify myocardial infarction on nonenhanced cardiac cine MRI achieved good diagnostic performance for detecting chronic myocardial infarction (per-segment sensitivity, 90%; specificity, 99%; area under the receiver operating characteristic curve, 0.94).
n There was no difference between the area of chronic MI detected on nonenhanced cardiac cine MRI and ground truth defined by expert manual segmentation of late gadolinium enhancement MRI (per-patient myocardial infarction area, 6.2 cm 2 6 2.8 vs 5.5 cm 2 6 2.3, respectively; P = .27). heart disease, pulmonary heart disease, myocarditis, infective endocarditis, and pericarditis; control patients and patients with poor image quality due to arrhythmia were also excluded. A subset of these (104 patients with chronic MI and 10 control patients) was included in a previous preliminary study (20).
LGE MRI was performed in the same orientations and with the same section thickness using a two-dimensional segmented fast low-angle shot phase-sensitive inversion recovery sequence (4.09/1.56; field of view, 284 3 350 mm 2 ; matrix, 163 3 256) 10 minutes after intravenous injection of gadolinium-based contrast agent (Magnevist, 0.2 mmol/kg; Bayer Healthcare, Berlin, Germany).
Full details of both protocols are given in Appendix E1 (online).

Ground Truth Definition
Data standardization and left ventricle localization.-Images from nonenhanced cardiac cine and LGE MRI were cropped au-tomatically into 64 3 64 pixels (pixel size, 1.46 3 1.46 mm 2 ), which included the full left ventricle area. Sections beyond the most base and apex regions were excluded manually.
Endocardial and epicardial contours delineation.-Endocardial and epicardial contours were manually delineated on the LGE MRI by a radiologist (N.Z., with 7 years of experience in cardiovascular MRI) after rigid registration (21) of the end-diastole phase of the cine and LGE images. Following visual inspection and assessment by mutual information, any residual registration errors were corrected using a diffeomorphic image registration technique (22) with parameterized deformation fields.
MI delineation.-MI was manually outlined on the LGE images by the same radiologist (N.Z.) after appropriate setting of the display window level and width. Microvascular obstructions were included in the MI regions. All manual segmentations (epicardial and endocardial contours and MI) were reviewed by another expert (L.X., with 10 years of experience in cardiovascular MRI) and in cases of disagreement, a consensus was reached. The MI area percentage (23) was calculated as (MI pixels/left ventricle myocardium pixels) 3 100%.  In the training data set (80% of study participants), a supervised deep learning framework was developed to locate the left ventricle (LV), to explore local and global motion patterns, and to compare with regions of MI manually outlined in LGE MRI. Predictive performance of the deep learning was assessed by using a further independent testing data set (20% of study participants). CMRI = cardiac MRI, ROI = region of interest, RNN = recurrent neural network.

Deep Learning
Our deep learning model extracts representative local and global motion features in nonenhanced cardiac cine MRI and relates them to LGE images (details in Appendix E1 [on-excluded (24). The MI area in each segment was calculated and each segment was categorized as having no MI, transmural MI (MI area . 50% of the myocardium), or subendocardial MI (MI area  50% of the myocardium) (25). segmentation, which has stored the version we used to achieve the current reported results.

Experimental Settings
The performance of our trained deep learning model was evaluated using independent testing, that is, a data set was not used for model development (external validation as mentioned in Park et al [28]). The 299 participants were randomly divided 80:20 into training data sets (169 patients with chronic MI, 69 control patients) and independent testing data sets (43 patients with chronic MI, 18 control patients). Basal, midcavity, and apical sections were analyzed in each participant, resulting in a total of 3808 segments for training and 976 segments for independent testing.
In addition, 10-fold cross-validation (29,30) on the whole data sets (with 299 participants) was performed to further confirm the effectiveness of our proposed deep learning model (details and secondary results are shown in Appendix E1 [online]). line]). Once the model is trained, predictions of MI location, size, and transmurality can be made without LGE images.
The deep learning framework consists of (Fig 2): (a) a localization deep network for detecting the left ventricle; (b) a motion feature extraction component incorporating local motion features extracted from a recurrent neural network and global motion features derived using an advanced optical flow method; and (c) a fully connected discriminative network (26) that distinguishes MI from normal myocardium. The deep motion networks output a probability map, and a threshold of 0.5 was used to create the final binary segmentation. The MI determined from nonenhanced cine versus from LGE MRI were compared with Dice scores (27). Analyses were performed using the full 25-phase cine data set and using a single end-diastolic frame, the latter having significant time savings.
Our implementation is open source and it is accessible at https://github.com/xuchenchuzw/MI-Segmentation#mi-

Computational Time
The parameters of our deep learning implementation are summarized in Figure E1 (online). The training time of our deep learning on the entire 238 data sets is 373 minutes (~1.6 minutes per data set). In the test phase, the computational time is 191 seconds for one MRI data set and about 8 seconds for one MRI section.

Independent Testing on Single End-Diastolic Phase and Full Nonenhanced Cardiac Cine Images
The quantitative results of using our deep learning framework on single end-diastolic and full nonenhanced cardiac cine image data sets are summarized in Tables 2 and 3. As expected, the

Statistical Analysis
Statistical analysis was performed using SPSS 23.0 (SPSS, Chicago, Ill). Independent t tests and x 2 tests were used to compare differences between two groups for continuous and dichotomous variables, respectively. Using manually delineated LGE images as the ground truth, sensitivity and specificity of the delineated MI derived from nonenhanced cardiac cine images with our deep learning framework were calculated. We also performed analysis of the area under the receiver operating characteristic curve (AUC).
MI area and MI area percentage at the segmental level were normally distributed (Kolmogorov-Smirnoff test) and differences from ground truth were assessed using paired t testing, Pearson correlation, and Bland-Altman analyses. Differences between count variables (per segment, per section, and per patient) were assessed by using the McNemar test. A two-sided P value less than .05 was considered to indicate a statistically significant difference.

Study Population Characteristics
No significant differences were found in weight and height between patients with chronic MI and control patients (training data sets, P = .82 and P = .23; independent testing data sets, P = .26 and P = .58). Table 1 and Table E1 (online) show that men were more common in the chronic MI cohort than in the control co-  performance of our deep learning on full nonenhanced cardiac cine data with 25 phases is significantly better than that obtained using a single end-diastolic image of the nonenhanced cardiac cine data. AUC analysis also shows that using the full nonenhanced cardiac cine image data set yields an overall AUC of 0.94 compared with an AUC of 0.58 for the single-phase data. The rationale for single-phase analysis is that it would be significantly faster to achieve. However, as shown in Tables 2 and 3, the results are substantially poorer and this was not pursued further.
By using the full nonenhanced cardiac cine image data, the overall sensitivity and specificity for the detection of MI segments were almost all higher than 90%, with the exceptions being sensitivity for detection of subendocardial MI ( Results were similar for basal, midcavity, and apical sections. There were no MI segments found in control patients. In addition, the Dice score between the MI segmentation from nonenhanced cardiac cine MRI and the ground truth segmentation in the independent testing data sets was 86.1% 6 5.7. Figure 3 shows three examples of the MI delineation results using the full nonenhanced cardiac cine image data sets compared with the manually segmented MI on LGE MRI. The MI transmurality is visualized as the predictive probabilities output from our deep learning framework, with probability greater than 0.5 being transmural and 0.5 or less being subendocardial MI. In so doing, we have a probability map to visualize the MI transmurality. Figure 4 shows AUC analyses for total, transmural, and subendocardial MI segments in basal, midcavity, and apical sections. All AUC values are 0.87 results in Figure 6 also show strong correlations of the MI area and MI area percentage (correlation coefficients, 0.88 for MI area and 0.89 for MI area percentage) measured from nonenhanced cardiac cine MRI and the manual delineated ground truth from the LGE MRI.

Discussion
In our study, we developed a fully automatic deep learning framework to detect chronic myocardial infarction (MI) in nonenhanced cardiac cine images based on extracted motion features. Using an independent testing data set, the Dice score (86.1% 6 5.7) and correlations (per-patient MI area, 6.2 cm 2 6 2.8 vs 5.5 cm 2   or greater, which shows the robustness of the developed deep learning framework. Table 4 summarizes the comparative results between our deep learning framework and LGE in the independent testing data sets (per segment) for all MIs and for transmural and subendocardial MI subgroups, with no significant differences. Table 5 summarizes the per-patient and per-section results, respectively. Ten-fold cross-validation showed very similar persegment, per-section, and per-patient results (Tables E2 and  E3 in Appendix [online]).
Pearson correlation and Bland-Altman analyses for MI area and MI area percentage are shown in Figure 5  Our framework only requires nonenhanced cardiac cine images, which are routinely acquired as part of a cardiac examination for function assessment. Other approaches using nonenhanced cardiac cine MRI combined with tagging and/ or feature-tracking techniques have also differentiated established or chronic MI from healthy remote myocardium. However, while Ogawa et al (19) showed that MI presence and position could be assessed using both feature tracking and myocardial tagging, the sensitivity and specificity for detecting MI segments were low (feature tracking: sensitivity, 72%; specificity, 71%; tagging technique: sensitivity, 71%; specificity, 75%). Fent et al (34) also reported that feature tracking could identify prior MI but the AUC was low (0.66 [95% CI: 0.54, 0.79], P = .012).
The feasibility of T 1r cardiac MRI (35,36) for nonenhanced detection of chronic MI has also been shown but the contrast-to-noise ratio between healthy tissue and MI is low and further sequence developments are required. More recently, texture analysis has been investigated for detecting subacute and chronic MI from nonenhanced cardiac cine; sensitivity of 86% and specificity of 82% was obtained by Bessler et al (37) and overall AUC of 0.85 was achieved by Larroza et al (38).
Our study has a number of limitations: (a) It is a proof-ofconcept study using retrospective data from a single vendor presence, position, transmurality, and size of chronic MI without requiring additional information from LGE images. Poorer overall sensitivity was obtained for subendocardial MIs (36 of 44 [81.8%]; 95% CI: 66.8%, 91.3%) compared with transmural MIs (78 of 83 [94.0%]; 95% CI: 85.8%, 97.7%) (P = .03), which may be attributed to their higher circumferential strain (19). The major contribution of our study is that chronic MI can be diagnosed from routinely acquired nonenhanced cardiac cine images.
Our study benefits from the fast development of deep learning techniques, which have demonstrated superior performance in medical image analysis by leveraging available big data (32,33). Previous studies have shown that some subtle changes in medical images, for example, progression of abnormality, can be distinguished with higher precision and sensitivity via deep neural networks than human visual inspection (32,33). Our study shows that myocardial wall motion may be used to accurately predict chronic MI area via deep learning. Overall, the deep learning approach achieved high sensitivity, specificity, and AUC for chronic MI detection. Compared with previous motion feature tracking methods, our framework (based on a recurrent neural network) can compress local motion distribution while extracting the global motion field from region of interest time series and generate a dense motion field to comprehensively characterize both local and global motions. Bland-Altman analysis (middle) of per-section analysis show that the myocardial infarction (MI) delineated by our deep learning on nonenhanced cardiac cine MRI is in accordance with the ground truth MI regions segmented on late gadolinium enhancement (LGE) MRI (independent testing data set). Segmentation results (right) are shown for a well-correlated section (Sample 1) and for a less well-correlated section (Sample 2). Dotted yellow line = ground truth manual segmentation from LGE images, green shaded region = MI detected from deep learning on nonenhanced cardiac cine images. and single center. (b) Our independent testing data set was small, consisting of 43 patients with chronic MI and 18 control patients (20% of all patients). (c) Our ground truth endocardium, epicardium, and MI delineations were performed manually by a single expert due to limited resources. These were then reviewed by a second expert who either ratified the first expert's segmentation or made minor modifications (by consensus following joint discussion). As such, we are unable to provide interrater agreement. It should also be noted that for our study, any microvascular obstructions were included in the MI regions, although this may affect the motion features (39). (d) For the single-phase method, we only tested an end-diastolic phase but an end-systolic phase may have performed better. However, this is not due to any inherent limitations in the methodology itself but reflects the fact that the LGE data that we had available to train the model was acquired in end diastole. (e) We have not assessed how the number of cardiac phases in the cine study affects diagnostic accuracy of the technique.
In conclusion, a robust deep learning framework for using nonenhanced cardiac cine MRI to infer the likely location, extent, and transmurality of myocardial infarction (MI) has been described, which can be readily expanded to future prospective studies. Future larger-scale studies with data from multiple sites are required for a full validation of our deep learning framework. These would also allow the accuracy of MI prediction to be determined for different myocardial segments with different motion characteristics. Further comparison with microvascular obstructions excluding data and texture analysis will be investigated in future work.