(A-15) The estimation of size and location of myocardial infarction using deep learning approach.

Introduction:: Myocardial infarction (MI), due to the blockage of coronary arteries, is a leading cause of death in the US and worldwide [1]. Cardiac scar size and location in MI can be identified using cine cardiac magnetic resonance imaging (CMR) with late gadolinium enhancement (LGE) [2]. LGE imaging based on a gadolinium-based contrast agent identifies matured infarcted regions by bright image intensity. However, LGE has some limitations including accumulation of gadolinium in tissues and prolonged scanning time which limit its scope and safety in time-sensitive clinical applications. Furthermore, the gadolinium-based imaging process may be dangerous to chronic end-stage kidney patients and approximately 20% of the patients who suffer from acute MI also have chronic kidney disease [3]. And, perhaps more importantly, LGE requires CMR which is not a feasible imaging modality at all health centers even in developed countries. Therefore, a less invasive and more feasible method to identify infarct region characteristics will assist in reducing undiagnosed MI cases. In this work, we developed a machine learning (ML) model to identify the location and extent of infarcted myocardium by employing cardiac strains that are feasibly obtainable from routine CMR scans and very possibly from echocardiography. We implemented a pipeline that takes raw CMR scans and calculates 4D (3D space plus time) strains. The 4D strains were used in the ML as inputs to predict the location and extent of infarct regions.


Materials and Methods:: The 3-D reconstructed heart geometry from CMR scans was truncated under the valve plane and meshed using quadratic tetrahedral elements as shown in Fig. 1a. Active and passive material properties were assigned to the finite element (FE) heart model, details are described in [4]. We developed a large database of semi-synthetically FE heart models covering numerous possibilities of MI region location and shape (Fig 1b) with variable stiffness of infarcted regions. The endocardium surface was pressurized with a pressure wave of a maximum amplitude of 7.5 mmHg. The average Green strains were computed at each segment according to American heart association (AHA) segmentation (Fig 1c) in circumferential, longitudinal, and radial directions. A 17×3 image was constructed using 17 AHA segments and three anatomical strains (circumferential, longitudinal, and radial) for each segment. This image was provided to the convolutional neural network (CNN) ML model to identify size and location of infarct regions (Fig 2). We built a binary array with the same size of the number of segments as labels in which 1 is infarct and 0 is non-infarct. Hundred examples with different sizes and locations of infarcts were generated to train the model. Leave-one-out cross-validation technique was utilized to reduce the variance with a limited number of examples. As the number of non-infarct regions was higher than the infarcted regions, the weights imbalance classification was applied, and weights were calculated and distributed at the time of training the CNN model.


Results, Conclusions, and Discussions:: Results and Discussion: The inverse modeling with estimated active and passive parameters and appropriate fiber architecture provided the end-systolic strains with the highest stiffness at the infarcted region (Fig. 1D). The ML model located the infarct by utilizing these cardiac end-systolic strains in the left ventricle (LV) and the performance was demonstrated in the form of a confusion matrix (Fig. 2a). Our approach accurately classified the infarcted regions from the normal tissues with 94.59% accuracy, 0.98, 0.96, and 0.97 sensitivity, precision, and F1-score, respectively. We validated our proposed model by comparing the ML predictions with ground-truth measurements by LGE imaging in MI human patients (Fig 2b). We used our in-house diffeomorphic image registration to calculate 4D strains (3D + time) using only short-axis slices. End-systolic strains at each segment of AHA segmentation (Fig. 1d) were provided to the ML model to classify the infarcted region which is compared with the LGE imaging to obtain the accuracy of our model. Our results suggest that ML model can delineate the MI using cardiac strains (from cine CMR or echocardiography) without using a contrast agent which shows the ability of the ML model to provide timely interpretations for clinical follow-up.

Conclusions: This study developed a CNN model to identify the location and extent of MI using cardiac strains obtained from routine CMR scans and potentially from echocardiography. By creating a large database of semi-synthetic FE heart models with varying MI region locations, shapes, and stiffness, we demonstrated the ability of the ML model to accurately classify infarcted regions from normal tissues with high sensitivity, precision, and F1-score. Our approach overcomes the limitations of gadolinium-based late gadolinium enhancement (LGE) imaging, which can be time-consuming, unsafe for certain patients, and not feasible at all healthcare centers.


Acknowledgements (Optional): : Acknowledgments: RA was supported by the NIH, under grant No. R00HL138288.


References (Optional): : References: [1] Bui, A. L. et al., Nature Reviews Cardiology, 8(1), pp.30-41, 2011. [2] Mendiola, EA et al., The FASEB Journal, 36, 2022. [3] Fox, C. S. et al., Circulation, 121(3), pp.357-365, 2010. [4] Mendiola, EA et al., Annals of Biomedical Engineering, 1-18, 2022.