Enabling Reliable Visual Detection of Chronic Myocardial Infarction with Native T1 Cardiac MRI Using Data-Driven Native Contrast Mapping

Abstract

Purpose

To investigate whether infarct-to-remote myocardial contrast can be optimized by replacing generic fitting algorithms used to obtain native T1 maps with a data-driven machine learning pixel-wise approach in chronic reperfused infarct in a canine model.

Materials and Methods

A controlled large animal model (24 canines, equal male and female animals) of chronic myocardial infarction with histologic evidence of heterogeneous infarct tissue composition was studied. Unsupervised clustering techniques using self-organizing maps and t-distributed stochastic neighbor embedding were used to analyze and visualize native T1-weighted pixel-intensity patterns. Deep neural network models were trained to map pixel-intensity patterns from native T1-weighted image series to corresponding pixels on late gadolinium enhancement (LGE) images, yielding visually enhanced noncontrast maps, a process referred to as data-driven native mapping (DNM). Pearson correlation coefficients and Bland-Altman analyses were used to compare findings from the DNM approach against standard T1 maps.

Results

Native T1-weighted images exhibited distinct pixel-intensity patterns between infarcted and remote territories. Granular pattern visualization revealed higher infarct-to-remote cluster separability with LGE labeling as compared with native T1 maps. Apparent contrast-to-noise ratio from DNM (mean, 15.01 ± 2.88 [SD]) was significantly different from native T1 maps (5.64 ± 1.58; P < .001) but similar to LGE contrast-to-noise ratio (15.51 ± 2.43; P = .40). Infarcted areas based on LGE were more strongly correlated with DNM compared with native T1 maps (R² = 0.71 for native T1 maps vs LGE; R² = 0.85 for DNM vs LGE; P < .001).

Conclusion

Native T1-weighted pixels carry information that can be extracted with the proposed DNM approach to maximize image contrast between infarct and remote territories for enhanced visualization of chronic infarct territories.

Keywords: Chronic Myocardial Infarction, Cardiac MRI, Data-Driven Native Contrast Mapping

Supplemental material is available for this article.

© RSNA, 2024

Summary

A data-driven approach was developed that exploits the unique T1-weighted MRI signal intensity patterns between infarct and remote myocardium to generate noncontrast images of chronic myocardial infarction, resulting in similar contrast-to-noise ratio compared with late gadolinium enhancement MRI.

Key Points

• ■ Native T1-weighted images acquired from canines with chronic myocardial infarction contain previously unrecognized signal characteristics in remote and infarct regions which can be used to enable generation of image contrast that is fundamentally different from what is available from T1 maps.

• ■ The proposed data-driven approach permits detection of chronic infarct territories without contrast agents, resulting in an image contrast-to-noise ratio (mean, 15.01 ± 2.88 [SD]) that was similar to that of the reference standard contrast-enhanced images (late gadolinium enhancement) (15.51 ± 2.43; P = .40) but vastly higher than that of native T1 maps (5.64 ± 1.58; P < .001).

Introduction

Postinfarction heart failure, a major health issue affecting millions in the United States (1,2), requires knowledge of presence and characteristics (location, size, and transmurality) of chronic myocardial infarction (MI) for effective treatment. American Heart Association guidelines recommend late gadolinium enhancement (LGE) cardiac MRI for obtaining this information (3,4). Although the requisite gadolinium-based contrast agents for LGE have evolved in safety over the past several years, their use in patients with MI can present some workflow challenges as all non–contrast-enhanced acquisitions need to be completed prior to LGE. The ability to obtain infarct characteristics without LGE enables tailoring of acquisitions to clinical need (particularly with respect to how and when specific acquisitions are performed) and a potential reduction in use and cost of contrast media (especially within the same examination) and repeat examinations. Some noncontrast alternatives to LGE have been proposed but have significantly lower contrast-to-noise ratio (CNR) than LGE (5–9). Recent studies have shown that native T1 maps combined with semiquantitative analysis can accurately delineate chronic MI (10–12). However, the visual detection sensitivity (~60%) of this approach is weak compared with LGE (~96%) (12), complicating the automatic segmentation of MI regions. Furthermore, manual and semiautomated methods often require extensive effort and expertise, underscoring the need for substantial improvement in CNR for native T1 imaging to be a viable clinical tool.

Native T1 mapping presumes that each voxel consists of homogeneous tissue (13,14), but the technique may be flawed given that chronic MI territories show a high degree of heterogeneity due to potential fat infiltration, iron deposition, surviving bundles, and fibrotic tissues, which affects spin behavior and contrast (15,16). A conventional T1 fitting process further reduces image contrast by forcing equal weighting across T1-weighted images over a wide range of inversion times (TIs) to ensure broad sensitivity to T1, potentially leading to the loss of valuable information present on T1-weighted images. To address these issues, the current study tested the hypothesis that suboptimal image contrast for visual detection of chronic MI with native T1 mapping could be improved by optimally utilizing T1-weighted images to increase CNR between remote and affected myocardium.

Materials and Methods

Study Design and Animal Model

We tested this hypothesis using highly aligned native T1-weighted and LGE images from a large animal model of chronic MI with variable tissue composition within the MI territories. MRI was acquired 6 months following intervention, where MI is considered chronic after 8 weeks in canines (17). Twenty-four canines (age, >24 months; 12 male and 12 female animals; strain, hound; source, Oak Hills Genetics) were studied. We examined voxel intensity patterns via unsupervised machine learning pattern clustering and visualization algorithms and trained a supervised deep neural network (DNN) for pattern modeling using a data set of over 100 000 pixel-wise intensity patterns (PIPs) of infarct and remote regions.

Quality Control and MRI Preprocessing

Breath-held native T1-weighted and LGE images covering the whole left ventricle in short-axis orientation were acquired from the 24 canines (details on animal preparation, histologic evaluation, and MRI are located in Appendix S1). Animal experiments performed in this study were approved by our institutional animal care and use committee. Eight T1-weighted images were acquired with TIs between 120 and 2500 msec. We selected the native T1-weighted image series, native T1 maps, and LGE images corresponding to sections positive for MI with similar TIs. For each section, four images from the native T1-weighted image series were selected, with TIs of approximately 200, 1000, 1750, and 2500 msec, respectively. Hereinafter, we refer to a set of 4 pixels corresponding to the same location on a native T1-weighted image series as a PIP. Infarcted regions, remote regions, artifactual regions (primarily from off-resonance effects at the heart-lung interface reflected by the balanced steady-state free precession readout at 3 T), and the blood pool were separately labeled by an expert reader (Y.C., radiologist with over 5 years of experience) as the reference standard for both LGE and native T1. Consistent with previous studies, the “mean + 5 SD” criterion was used to identify the infarct regions on LGE and native T1 maps (11). This resulted in ~132 000 PIP samples corresponding to infarct and remote regions for generating the PIP training data set for our DNN.

Deep Learning Modeling

To test whether optimal visualization of native T1 differences between infarcted and remote myocardium is best gleaned from native T1-weighted images than from native T1 maps, a deep learning (DL) model (18–20) was trained to fit PIPs from T1-weighted images into corresponding pixels from LGE. Given our pixel-based approach, no texture information was available to the model. Thus, the risk of data leakage from adjacent sections was mitigated. The performance of the DL model was evaluated using n-fold cross-validation (21), where n is the number of sections in the data set. PIPs corresponding to the section being tested were left out from training, and PIPs of all remaining n–1 sections were used for training, specifically PIPs corresponding to infarct and remote regions. This process was repeated for each section in the data set. Validating PIPs of a whole section at a time enables us to reconstruct the section from the left-out PIPs to perform subsequent section-based tests. In what hereinafter is referred to as data-driven native mapping (DNM), images of the myocardium were reconstructed using the pixel intensities determined by the DL mapping model. We note that the varying number of sections per subject in the data set complicates cross-validation at the animal level due to inconsistencies in training data volume and diversity, especially for animals that have multiple sections in the data set. However, the nature of our model’s input, which has access to only information retrieved from 1 pixel per T1-weighted image at a time (4 total), inherently excludes exposure to animal-specific characteristics. Nevertheless, for animals represented by a single section in the data set, our cross-validation is effectively interanimal. Thus, although our study’s cross-validation is primarily section-based, it occasionally aligns with animal-level validation. Figure 1 illustrates the DL modeling pipeline used to generate DNM images. The DL model architecture and training details can be found in Figure S1.

Figure 1:

Generating data-driven native mapping (DNM) with a deep learning modeling pipeline. Pixel intensities extracted from spatially aligned pixels over a series of T1-weighted images are collected to form the pixel intensity patterns (PIP). In the training phase, feature vectors derived from PIPs are used as inputs to a deep learning model, which is trained to map the PIPs to the corresponding late gadolinium enhancement (LGE) pixels’ intensities. In the testing phase, the trained deep learning model is used to map PIPs of a T1-weighted image series that was not used during training to an intensity level that resembles a corresponding LGE pixel. The deep learning model is used to map all myocardial PIPs which are then reconstructed into an LGE-like image that we refer to as the DNM image. The multistage training (MST) architecture and details are provided in Appendix S1.

Visualization of T1-weighted PIPs

Distinctive behaviors exhibited by PIPs corresponding to different categories of pixels from T1-weighted images, both native and contrast-enhanced (LGE), on an aggregate level and granular level were visually observed. Aggregate visualization (overall ability to differentiate PIP) was assessed by comparing the mean PIP of various pixel categories corresponding to the infarct, remote, artifact, and blood pool regions.

For a more exhaustive assessment of the ability of PIPs to differentiate between pixels corresponding to infarcted and remote regions at a granular level, an automated clustering technique was used. Each individual PIP was represented as a projection to one point on a two-dimensional graph such that the spatial location of the point is dependent on pattern similarity between PIPs as determined by the clustering algorithm. This was performed with a self-organizing map algorithm (22) to cluster PIPs corresponding to infarct and remote regions into four subcategories each and then randomly picking one infarct and one remote subcategory. t-distributed stochastic neighbor embedding (23) was used to generate the scatterplots.

Determination of Infarct Area

Infarct regions were automatically segmented using parameter sweep to determine the optimal windowing and intensity threshold and segmented manually by three expert cardiac MRI readers (K.V., imaging cardiologist with over 10 years of experience; A.K., imaging cardiologist with over 15 years of experience; A.H., imaging cardiologist with over 15 years of experience; details in Appendix S1). Infarct area from each imaging section was determined using a 96-segment myocardium model with 96 angular segments. Infarct size was calculated as a percentage of infarcted segments to adjust for myocardium size variability. A segment with over 15% of infarcted pixels (as defined using the mean + 5 SD criterion; see above) was considered as an infarcted segment. Otherwise, the segment was considered as a noninfarcted segment. This criterion was applied to both the LGE and native T1-based images.

Statistical Analysis

Statistical analyses were performed to compare outcomes from DNM, standard T1 maps, and LGE. Pearson correlation coefficient was used to assess the linear relationships, expressed as R² values. Distribution differences in contrast values between methods were analyzed using independent samples t tests. Agreement between DNM and other imaging methods was evaluated using Bland-Altman analysis which included calculating mean bias and limits of agreement. These analyses were conducted using MATLAB (version 2022a; MathWorks) with a significance level set at P < .05. Results indicating significant differences or agreements were noted based on this threshold.

Results

PIP Analysis of T1-weighted Images

PIPs from various regions of the left ventricle (shown in Fig 2) revealed a distinctive behavior of PIPs corresponding to MI regions relative to remote and artifact pixels, as well as blood pool pixels. Note that while the intensities corresponding to a single TI overlapped across different pixel types, PIPs comprised of relationships between pixels from multiple TIs exhibited robust discrimination of different pixel types. In other words, pixel patterns correlated with their corresponding tissue type.

Figure 2:

Pixel-intensity patterns determined from native T1-weighted signal intensities over multiple inversion times (TI) between infarct and other territories (blood pool, remote, and artifact) are distinct. (A–C) Graphs show the mean T1-weighted pixel-intensity patterns (solid lines) ± SD (dotted lines) for remote, blood pool, and artifact regions, respectively, as compared with infarct regions. (D) Schematic of an example image with different left ventricular regions: blood pool, infarct, remote, and artifact. Note that while the intensities corresponding to a single inversion time overlap among different types of regions, the normalized signal intensity relations over multiple inversion times yield distinct pixel intensity patterns for the different regions.

Individual PIP behavior from the infarct and remote categories analyzed with automated clustering is visualized in Figure 3. This visualization revealed that labeling the scatterplot using corresponding LGE pixels yielded better separability between infarct and remote myocardium pixels than labeling the same scatterplot using corresponding pixels from native T1 maps.

Figure 3:

Granular visualization of T1-weighted pixel intensity patterns reveals higher separability between infarct and remote clusters when labeled using late gadolinium enhancement (LGE) as compared with native T1 maps. (A, B) Cluster plots show unsupervised pixel-wise T1-weighted clustering performed on T1-weighted pixel intensity patterns (PIPs) from infarcted and remote regions using self-organizing maps and t-distributed stochastic neighbor embedding. A PIP is represented by a four-dimensional vector consisting of one myocardium pixel intensity on four spatially aligned T1-weighted images (one for each inversion time; see Fig 2). Each PIP is projected into one dot with a spatial location based on the similarity to other PIPs, as determined by the clustering algorithm. Similar PIPs are projected close to each other and dissimilar PIPs are projected far from each other. The dots where each PIP is projected is given a color based on the corresponding pixel type (blue, infarct region; red, remote region) in an (A) aligned LGE image and an (B) aligned native T1 map based on the manual expert segmentation. Note that when the PIP projections are colored based on corresponding LGE pixel types, they demonstrate highly separable clusters corresponding to infarct and remote pixels. On the other hand, when the PIP projections are colored based on corresponding native T1 pixel types, a weaker separation between infarct and remote clusters is observed. The arrows in B point to locations of disagreement between native T1 and LGE labeling, where blue arrows indicate false positives and black arrows indicate false negatives as compared with LGE. This observation aligns with the lower contrast differentiation on native T1 images. LGE labeling underscores the superior contrast potential of T1-weighted images, suggesting that traditional native T1 mapping does not effectively maximize contrast between infarcted and remote pixels.

DNM Image Evaluation

Unsupervised clustering analysis revealed that the potential for visual separability of T1-weighted PIPs for improved delineation of infarcted pixels is beyond what generic native T1 mapping can offer. Observations are summarized below.

Representative cases: visualization of infarct area.—Two representative cases of hyperintense and hypointense examples are shown in Figure 4. Case A is an example of a native T1 map with a hypointense infarcted region due to the presence of iron, which is improved by DNM, where all infarcted pixels (including those with iron-rich regions) are hyperintense as typically observed at LGE. Case B is an example where native T1 corresponds well with LGE, and DNM further improves the image contrast, making it easier to visualize and locate the infarct areas similar to LGE. Corresponding T1-weighted images of these cases are shown in Figure S2.

Figure 4:

Images show representative cases. Data-driven native mapping (DNM) markedly increases image contrast between infarct and remote myocardium on native T1 maps, allowing for robust visualization of chronic myocardial infarction (MI) territories (independent of iron status) similar to LGE but without contrast agents. Case A shows a set of spatially aligned short-axis (native T1, late gadolinium enhancement [LGE], and DNM) images of the heart from a canine with chronic MI and evidence of iron deposition (Fig S2 shows corresponding T1-weighted images). (A, B) Native T1 map shows hypointense zone (red arrows) within the MI zone (A) but not on the LGE image (B). (C) DNM model applied to the native T1-weighted images overcomes the hypointense zones identified in native T1 map and amplifies the image contrast between infarct and remote myocardium compared with the native T1 maps. (D) Histologic evidence supports chronic nature of MI (Masson trichrome–stained [original magnification, ×10] section indicates extensive replacement fibrosis). (E) Iron deposition within MI (Perls stained [original magnification, ×100] section; arrows point to iron particulates within the MI zone). Case B shows a similar set of short-axis images from another animal with chronic MI but without iron deposition. (F, G) Unlike in case A, the corresponding native T1 map (F) shows no hypointense zones within the MI territory, and the hyperintense zones are visually consistent with LGE (G). (H) DNM model amplifies the image contrast on the native T1 maps, permitting robust visualization of chronic MI territories without contrast agents. Masson trichome–stained section (I) and Perls-stained section (J) images show histologic evidence of chronic MI and absence of iron.

Aggregate data: visualization of infarct area.—Grouped results are shown in Figure 5. Consistent with representative data, the collective data show significant improvement in visualization of infarct territory based on the DNM approach compared with native T1 mapping. This is evidenced by the respective CNRs (mean, 5.64 ± 1.58 [SD] [native T1] vs 15.01 ± 2.88 [DNM]; P < .001), and the CNR of DNM was comparable to that of LGE (15.51 ± 2.43; P = .40). In addition to improving CNR between infarct to remote contrast in general, the DL model was also able to detect hypointense regions, which are identified as high intensity pixels on DNM images (Fig 4).

Figure 5:

Multifold amplification of contrast-to-noise ratio (CNR) between infarct and remote territories of native T1 maps by data-driven native mapping (DNM) leads to CNR levels similar to late gadolinium enhancement (LGE). (A) Bar graph shows mean CNR comparison between native T1, LGE, and DNM on infarcted sections. Note that when hypointense cases are excluded, the native T1 mean CNR is 6.85. (B–D) Graphs show the mean pixel intensities from two myocardial regions (infarct and remote) for each test case. The horizontal axis represents the mean pixel intensity, and the vertical axis represents the sample number, where each sample refers to an individual section. Blue dots correspond to infarct regions, and red dots correspond to remote regions. Values were normalized by the mean intensity of the two regions. Significant contrast improvement between infarct and remote regions is observed with the DNM as compared with the traditional native T1 maps. Red lines on the native T1 map panel correspond to sections with a majority of hypointense pixels, where the mean pixel intensity in the infarct region is lower than the remote region. Red arrows point to the corresponding LGE and DNM sections. Cases A and B (shown in Fig 4) are examples of sections with hypointense and hyperintense regions, respectively.

Quantification of infarct area: native T1 versus DNM versus LGE.—Figure 6 presents the comparison of infarct area estimated with native T1, DNM, and LGE. Correlation and Bland-Altman analysis plots of fully automated segmentation results are presented in Figure 6A–6C and Figure 6D–6F, respectively. Here, infarct sizes were determined automatically, where the windowing and the intensity threshold were optimized algorithmically. Since our dataset included some samples with hypointense native T1 infarcts, native T1 had a lower correlation with LGE (R² = 0.48). Corresponding correlation and Bland-Altman analysis plots from experts’ manual segmentations are presented in Figure 6G–6I and Figure 6J–6L, respectively, and show that experts can recognize the infarcted regions on native T1 images despite being hypointense. Here, the correlation between native T1 and LGE improved significantly (R² = 0.71). DNM images had significantly higher correlation in both scenarios (R² = 0.73 [automated] and R² = 0.85 [manual]; P < .001). Bland-Altman plots further corroborate the improved agreement between DNM and LGE as compared with that between native T1 and LGE. The Bland-Altman bias (–0.016 ± 0.028) for manual segmentation was lower as compared with the automated scenario (bias = –0.04 ± 0.042), indicating the qualitative visual improvement enabled by DNM and its close similarity to LGE.

Figure 6:

Bland-Altman and correlation plots show that fully automated quantification and expert reads of infarct size from data-driven native mapping (DNM) are more in line with late gadolinium enhancement (LGE) compared with standard native T1 maps. (A, B) Correlation plots compare infarcts sizes obtained from DNM and native T1 images to infarct sizes obtained from LGE images, respectively. (C) Correlation plot shows comparison of infarct sizes from DNM images to infarct sizes obtained from native T1 images for completeness. (D–F) Bland-Altman plots corresponding to A–C, where the lighter lines delineate the limits of agreement. The infarct sizes in A–F were determined automatically, where the windowing and the intensity threshold are optimized algorithmically. The low performance of native T1 is due to the hypointense cases that cannot be detected with the automated approach. (G, H) Correlation plots comparing infarcts sizes obtained from DNM and native T1 images to infarct sizes obtained from LGE images, respectively. (I) Correlation plot shows comparison of infarct sizes from DNM images to infarct sizes obtained from native T1 images are for completeness. (J–L) Bland-Altman plots corresponding to G–I. With G–L, infarct sizes of cases corresponding to animals with dampened infarct intensity on native T1 maps were manually annotated by three experts. Each expert annotated the infarcts independently, and the results were combined using majority voting (ie, pixels that were labeled as infarcted by two experts were considered infarct pixels).

Discussion

Our study demonstrates that, while standard native T1 mapping can identify chronic MI territories, at times there are tissue-specific components within the chronic infarct territories that can compromise visualization of infarcted regions. Unsupervised pixel-wise clustering and pattern analysis of native T1-weighted image series revealed the potential of optimizing infarcted pixel contrast relative to remote regions beyond that of standard native T1 mapping. Machine learning–based data-driven mapping using DNNs yielded images that more closely resemble LGE, with significant improvement in infarct contrast and visual conspicuity. Our analysis further indicates the potential of data-driven native T1-weighted mapping to generate models that are optimized to identify different types of substrates other than just fibrotic regions, which can help to further characterize the chronic MI zone without contrast agents.

For pattern visualization, we employed unsupervised machine learning clustering and visualization algorithms. Qualitative visual assessment provides a conceptual framework and supports the notion that with the appropriate model applied to the T1-weighted image series, it is possible to visually discern infarct territories from native T1-weighted images that closely match LGE images. For pattern modeling, a DNN was trained on over 100 000 PIPs of infarct and remote regions. The resultant data-driven mapping model closely matched LGE images. By focusing on a pixel-level approach, we prevent the DNN from depending on texture information, allowing a direct and fair comparison against the standard T1 mapping model.

Our results show that DNM images had significantly higher correlation with LGE than standard T1 mapping in both automated and manual segmentation. The performance improvement in automated segmentation demonstrates the DL model’s ability to amplify the infarct contrast regardless of presence of substrates that decrease T1 within the infarcted myocardium. The performance improvement in manual segmentation demonstrates the qualitative visual improvement that helped the experts segment infarcted regions more accurately. The improved correlation in the manual DNM segmentation scenario indicates that the experts can utilize additional information to identify infarcted regions, such as local texture patterns and spatial image features.

Our findings show that pixel-level information from T1-weighted images can be effectively used by data-driven mapping models to maximize image contrast, thereby enhancing visualization of chronic MI territories and significantly improving CNR. This approach also lays the groundwork for future inclusion of texture information, leading to an optimized T1-weighted mapping model for characterizing chronic MI without contrast agents. Another potential advantage of our DL-based method is the increased robustness to issues arising from beat-to-beat variability (eg, due to arrhythmias). However, this remains to be investigated in future work.

Our study had some limitations. Pixel-wise modeling has its strengths and weaknesses. Unlike image-based DL approaches that have been recently proposed to generate synthetic LGE images (24,25), pixel-wise modeling ensures that the model is focused on the interrelationships between native T1-weighted pixel intensities and is not biased toward texture information, which can lead to overfitting. Furthermore, it can enable us to improve our understanding of the contribution of individual T1-weighted images and the interactions between them, which will lead to designing optimal acquisition protocols tuned to TI ranges that maximize infarct to remote myocardial CNR. On the other hand, the inability to capture texture information and spatial features like image-based DL models poses a limitation on the segmentation performance, and the dependency of pixel-wise modeling on the interrelationships between native T1-weighted pixel intensities makes it sensitive to large deviations from the range of TI values used in training. Accordingly, this is a proof-of-concept study since we had to limit our model to a modest range of TI values. Future work should incorporate a hybrid approach that combines both pixel-wise and image-wise DL models to further improve robustness, performance, and explainability of automated MI segmentation models with cardiac MRI performed without contrast agents. Future studies should also expand upon the findings of this study by using larger data sets to train a more generalizable model, further improving DNM correlation with LGE and eliminating the sensitivity to TI which will allow for the development of models that can be quickly translated into clinical practice. Future work with larger data sets will also enable the use of a fixed number of sections per animal and allow us to further explore cross-validation at the animal level. Further, we did not apply our approach to other noncontrast chronic MI cardiac MRI approaches, including other T1 methods. Additional studies are needed to address this limitation. Moreover, while the approach has the potential for detecting the peri-infarct zone, this was not investigated in this work. Finally, we also did not delve into the biophysical mechanisms contributing to the PIP differences, particularly spin exchange across various tissue compartments. Additional studies are needed to rigorously investigate whether the observed difference in PIP is affected by spin exchange between various compartments with or without gadolinium in the remote and infarct area.

In conclusion, this study represents a proof-of-concept, data-driven, noncontrast cardiac MRI approach for accurately delineating chronic MI territory using a series of native T1-weighted cardiac MRI data. In this work, we developed and tested the proposed approach in a clinically translatable large animal model with histologic validation that is essentially identical to that used in the development of LGE cardiac MRI. A clear next step would be to translate our approach under clinical settings using large data sets comprising of T1-weighted cardiac MRI data (from patients with chronic MI and healthy individuals) and training a human-specific model. This would require that the proposed model is tuned to the parameters used for clinical cardiac MRI acquisitions (such as spatial resolution, heart rate, etc). Further, while the current work is based on previous studies that advocate the use of 3-T systems, the utility of the proposed approach remains to be investigated with 1.5-T systems as well.