Modeling hIerarchy between Cardiac descriptors with MAChine learning – MIC-MAC

-

-

-

Journal articles:
1. Duchateau N, King A, De Craene M. Machine learning approaches for myocardial motion and deformation analysis. Frontiers in Cardiovascular Medicine 2020;6:190.

Conference articles:
1. Di Folco M, Guigui N, Clarysse P, Moceri P, Duchateau N. Investigation of the impact of normalization on the study of interactions between myocardial shape and deformation. Proc. Int. Conf. on Functional Imaging and Modeling of the Heart (FIMH), LNCS 2021. In press.
2. Mom K, Clarysse P, Duchateau N. Population-based personalization of geometric models of myocardial infarction. Proc. Int. Conf. on Functional Imaging and Modeling of the Heart (FIMH), LNCS 2021. In press.

Conference abstracts:
1. Duchateau N, Viallon M, Petrusca L, Clarysse P, Belle L, Croisille P. Pixel-wise statistical analysis of lesion patterns: a fresh look at immediate vs. delayed stenting of the Minimalist Immediate Mechanical Intervention approach (MIMI) in acute STEMI. Society for Cardiovascular Magnetic Resonance (SCMR) congress. 2021.

The recent breakthroughs in machine learning are changing the way we consider data collection and processing. In medical applications, critical risks arise for the exploration of large databases, where disease understanding and decision-making remain extremely challenging for computational methods. Representation learning and in particular unsupervised learning are suited for knowledge discovery and stratifying risk among medical populations, but they face complex data integration issues. The data descriptors are numerous, and may be high-dimensional and of heterogeneous types. Besides, their combination is not straightforward. Current methods consider all descriptors at the same level, which is highly critical in terms of efficiency and interpretability, and a strong limit for knowledge discovery.
MIC-MAC aims at substantially improving the analysis of medical imaging databases on these aspects. It first proposes to revisit the data integration approach, with novel methodological schemes that take into account an existing hierarchy in the input imaging data. Then, it will move this further by developing methods that explicitly learn such hierarchies from data. This learning will not only consider the gain in knowledge provided by new data descriptors, but also the costs related to the clinical and processing context. This approach is highly original and under-explored, although crucial in healthcare. Finally, specific efforts will be devoted to enhance clinical transfer and interpretation, with the development of practical software tools to navigate through complex data and results.
The project will be centered on cardiac applications, which represent the perfect field to boost its success given the advanced PI’s expertise within both the methodological and clinical communities. It plans the retrospective exploration of large existing imaging studies of heart failure patients, from widespread imaging protocols (magnetic resonance and echocardiography). It will therefore highly benefit from an ideal environment at the host institution, with advanced expertise on medical image processing workflows for ?cardiac applications, and the involvement of renowned external clinical partners with access to complementary large databases.
The project is structured into four workpackages: the first two are methodological workpackages that address novel aspects on the hierarchical integration of high-dimensional data (WP1) and the learning of this hierarchy from data while accounting for real-life costs (WP2). Then, WP3 is devoted to the clinical application and will address data pre-processing and software development aspects that are fundamental for the methodological workpackages. Finally, WP4 will serve for scientific and data management along the whole project.
MIC-MAC is a unique opportunity to strengthen the PI’s research both on the methodological side (the statistical analysis of multi-parametric high-dimensional data) and the applicative side (cardiac imaging), and develop his abilities to lead research. Revisiting the data integration schemes and considering the actual costs involved has high potential for impacting the ongoing data collection context, and the design of machine learning algorithms beyond the medical application. Naturally, impact will be maximal in healthcare, with strong efforts to benefit the clinical community with intuitive and interpretable methods.