(E-189) Estimation of volume of the right ventricle from the endocardium area of CMR slices using recurrent neural network.

Introduction:: The right ventricle (RV) is a critical component of the cardiovascular system and accurate assessment of RV volume is crucial for the diagnosis, management, and prognosis of various cardiac conditions, including heart failure, pulmonary hypertension, and congenital heart diseases [1]. Echocardiography and cardiac magnetic resonance (CMR) have been widely used for evaluating right ventricular function, however, these methods rely on geometric assumptions and can be subject to inter-observer and intra-observer variability, which may result in inaccurate and inconsistent measurements [2]. The volume estimation of RV from CMR and echocardiography slices remained labor-intensive and time-consuming, often leading to inconsistencies in RV volume estimation [3]. In this study, we propose a recurrent neural network approach for estimating right ventricular volume from the endocardial area of RV from the in-silico rat heart models then validated with the CMR slices. This method aims to overcome the limitations of manual RV volume calculations, enhancing the accuracy, consistency, and efficiency of the process.


Materials and Methods:: In this study, we utilized a total of 25 male rats. This group comprised nine Sprague-Dawley, nine Fischer-344, and seven Wistar Kyoto rats. The animals were classified into three categories: healthy rats (n=8), rats with pulmonary hypertension (PH; n=11), and rats with myocardial infarction (MI; n=6). After acquiring images of those rats, the total number of geometries was expanded to 250 geometries in which each geometry created ten additional geometries by keeping the epicardium the same and varying the size of the RV and left ventricle. The 3-D heart geometry was reconstructed from CMR scans (Fig. 1a, b) to prepare it for the finite element (FE) model. Active and passive properties were applied to the FE model [4] and endocardial areas of RV at different slices (Fig. 1c, d) is calculated as a result of applying a pressure wave of a maximum amplitude of 7.5 mmHg. RV endocardium areas at different time points between end-diastole to end-systole are provided to the recurrent neural network (RNN) to memorize the volume at the time steps to accurately delineate the volumes at all time steps.


Results, Conclusions, and Discussions:: Results and Discussion: Our study demonstrated significant findings in the assessment of RV volume using the proposed RNN approach. The RNN model successfully predicted the volumes of RV at all provided geometries of rats at all time points using endocardial areas of slices. Notably, the model exhibited improved accuracy and consistency compared to traditional manual methods, while also reducing the time and effort required for analysis.

Conclusions: In conclusion, this study highlighted the potential of machine learning in estimating RV volume from the endocardial area of CMR slices. The results demonstrated improved accuracy, consistency, and efficiency compared to traditional manual methods, which can significantly enhance the evaluation of RV function in a variety of cardiac conditions. Furthermore, the RNN approach has the potential to be integrated with other diagnostic tools, providing a comprehensive and efficient method for assessing cardiac function and structure.


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


References (Optional): : References: [1] Haddad F., et al. Circulation 117.11: 1436-1448., 2008. [2] Medvedofsky D., et al., Journal of the American Society of Echocardiography, 28(1), 1-39, 2015. [3] Petitjean, C., Medical Image Analysis, 15(2), 169-184., 2011. [4] Mendiola, EA et al., Annals of Biomedical Engineering, 1-18, 2022.