Automated Processing and Analysis of the Human Right Ventricle for the Detection of Pulmonary Hypertension

Cardiovascular health has been and continues to be one of the most important research topics in worldwide healthcare. Hypertension in particular has been a major concern for many years, and yet, the changes in heart function due to sustained pressure overload are still not well understood. For example, pulmonary hypertension (PH) is a deadly disease that is well known to considerably change the appearance and function of the heart, especially the right ventricle 1. However, there are no clear quantitative metrics relating to these changes in the heart that are available to physicians to accurately predict PH patient outcomes.
Previous research has shown the existence of a relationship between the shape of the right ventricle and the progression of PH 2. More specifically, prior work has relied upon describing the shape of the right ventricle endocardial surface (RVES) through harmonic mappings to the sphere, with dimensionality reduction through direct methods such as PCA/POD 1 or spherical harmonics 2. However, these methods required manual alignment and feature identification under supervision from a trained cardiologist, resulting in significant preprocessing expense and a reduced dataset to train classifiers on. These preprocessing challenges are the most significant limitation in preventing further investigation of this link between RVES shape and the state of PH. As such, the main objective of the proposed research is to develop a machine learning approach for automated image extraction and processing to evaluate patterns relating to the shape and function of the human heart related to the state of PH.
Several avenues are expected to be explored for integrating machine learning to substantially improve the efficiency and reliability of the process to analyse heart shape. Potential areas to explore include techniques such as Laplace-Beltrami Surface Mapping, which allow for automated detection of pole and dateline features, and have previously been demonstrated to create harmonic maps for structures within the brain 3. Another approach would be to avoid the need for manual alignment through the use of rotationally invariant features 4,5. Alternatively, with a much larger set of preprocessed data, more advanced neural network techniques such as autoencoders 6 or deep belief networks 7 could be used for dimensionality reduction. Direct classification through neural networks is also possible through convolutional networks applied to the surface of the harmonically mapped sphere 8 or the image data directly, or by fully connected networks applied to features detected through previously mentioned methods 9.
References
1 Wu, J. et al. Phys. Med. Biol. 57, 7905 (2012)
2 Wu, J. et al. Comput. Methods Biomech. Biomed. Eng. Imaging Vis. 4, 327-343 (2016)
3 Shi, Y. et al. in Medical Image Computing and Computer-Assisted Intervention (MICCAI) 2008 147-154 (Springer Berlin Heidelberg, 2008)
4 Kazhdan, M. et al. in Proceedings of the 2003 Eurographics/ACM SIGGRAPH Symposium on Geometry Processing 156-164 (2003)
5 Skibbe, H. et al. in 2009 IEEE 12th International Conference on Computer Vision Workshops, ICCV Workshops 2009 1863-1869 (2009)
6 Baldi, P. 37-49 (2012)
7 Hinton, G. E. et al. Neural Comput. 18, 1527-1554 (2006)
8 Maron, H. et al. {ACM} Trans. Graph. 36, 1-10 (2017)
9 LeCun, Y. et al. in 9-50 (Springer, Berlin, Heidelberg, 1998)