4D Cardiac MR Images

This project finished in 2013 – now continued by “Medilearn”

Goals of the project

Given a large database of cardiac images of patients stored with an expert diagnosis, we wish to automatically select the most similar cases to the cardiac images of a new patient.


This is important, for instance to try and estimate the optimal time for an operation when a Tetralogy of Fallot condition is diagnosed (see fig. 1).


We want to be able to index images based on e.g. the shape of the heart, the dynamics of the myocardium, or the presence of anomalies. This would require capturing the right level of shape, motion and appearance characteristics in a compact and efficient way so as to enable fast indexation and retrieval even for hundreds of 4D datasets. Using state of the art efficient machine learning techniques is also of paramount importance. The design of both the visual features and the classification algorithms have to be informed by medical experts so as to make sure the final system remains relevant from a clinical point of view.

Such a system would provide a strong innovation in cardiology, a learning tool for residents and young physicians, and a new tool to help physicians to assess a diagnosis and plan a therapy. The method could emphasize local motion and/or local shape singularities (e.g. septal flash (motion anomaly) vs. ventricle overload (shape anomaly), with the possibility to actually weight motion vs. shape features.

The MSRC­-Inria collaboration

Key points to succeed in developing such a system are the following:

  • Expertise in cardiac image processing, in order to extract automatically the boundaries of ventricular cavities (endocardium of left and right ventricles) from cardiac MR images, and outer boundaries of the heart (epicardium), as well as extraction of cardiac motion (dense non-­‐rigid motion of myocardium, accounting for quasi incompressibility constraints of tissues and isovolumetric phases of cardiac motion).
  • Expertise in indexation of large databases of complex images, in order to select the appropriate shape and motion invariant features, and to select an appropriate metric in feature space reflecting the “clinical distance” between different medical cases, as well as the development of efficient algorithmic solutions to beat the curse of dimensionality
  • Expertise in pediatric cardiology, with the constitution and access to a very large database of 3D + time cardiac images stored with an expert clinical diagnosis including the specificity of each case but also the therapy choice along with the corresponding outcome.


Our partnership

It is clear that the close partnership between INRIA Sophia Antipolis, Microsoft Research Cambridge, and King’s College London meets the required expertise of the three points above, and could lead to a pioneering advance in the field of computer aided clinical decision for pediatric cardiology and more generally to the field of content-­‐based retrieval of medical images.

The best strategy to implement this research agenda would be to rapidly hire a highly qualified PhD student, who would share his/her time between the 3 institutions, in order to rapidly develop expertise by leveraging the extensive knowledge of three participating laboratories. The exact repartition of the time would probably depend on the original qualification of the candidate, and on the progress made during the PhD research.