Automatic Classification of Cardiac Examinations after Contrast agent InjecTion – ACCECIT

Every day a large number of cardiac MR data are acquired, generating an increasing mass of data. The manual analysis of these medical images is a long and tedious task. Thus there is a need of reliable tools to automatically segment regions of interest and extract clinical parameters from these medical images. In this project, we will focus on late gadolinium enhancement (LGE) and T1 mapping cardiac MR images.
LGE imaging has been widely used for detection and assessment of myocardial scar and presence of fibrosis in cardiac MR imaging. It is based on the fact that there is a difference in signal intensity between normal and fibrotic myocardium. LGE is a very accurate technique to detect myocardial infarction (MI) or microvascular obstruction (MVO). Other myocardial diseases like myocarditis, amyloidosis, hypertrophic cardiomyopathy (HCM), dilated cardiomyopathy (DCM) also show different patterns of LGE underlining the usefulness of LGE in the diagnostic or the prognostic of these diseases.
LGE is a gold standard for the quantification of focal myocardial fibrosis, but in some cardiomyopathies the fibrosis is often diffuse and cannot be quantified on LGE images.

To overcome this problem, T1 mapping techniques have been developed to quantify diffuse myocardial fibrosis and to characterize tissues. By measuring myocardial and blood T1 before and after administration of contrast agent, the extracellular volume (ECV) fraction can be computed. For example, T1 mapping imaging can be a useful method to evaluate the severity of the fibrosis in HCM. Moreover, CMR-based assessment of ECV may have the potential to serve as a non-invasive tool for the quantification of diffuse myocardial fibrosis in order to monitor therapy response and aid risk stratification in different stages of DCM.

The aim of this project is to combine automatically information from LGE and T1 mapping images by developing a new artificial intelligence tool combining Bayesian statistics and deep learning to detect areas of fibrosis or abnormal tissues and to automatically classify the different pathologies and identify normal cases.

This goal is facing several problems:
- Images can be acquired on different sites with different scanners with different magnetic fields (1.5T or 3T), leading to heterogeneous data, thus we need a robust tool independent of the scanner used for the images acquisition.
- LGE images need to be automatically segmented whatever pathology the patient is suffering from.
- Heart contours need to be automatically detected on T1 mapping images whatever pathology to extract a robust and normalized T1 value.
- Information coming from the three series of data (LGE images, T1 mapping before and after contrast agent injection) need also to be combined for pathology classification.

The objectives of the project are multiple:
- Objective A: First we will collect and manage the database containing three series of data (LGE images, T1 mapping before and after contrast agent injection) for each studied pathology (myocardial infarction, DCM, HCM, myocarditis) as well as for normal cases.

- Objective B: We will develop deep learning tools to automatically segment heart contours, detect areas of fibrosis or abnormal tissues, extract physiological parameters and classify the examinations according to the pathology. To this aim, we will investigate Bayesian deep learning methods which combine Bayesian statistics with deep networks to obtain true network uncertainty estimates. Classical deep learning methods will also be tested to compare the results and to quantify the improvements of proposed method.

- Objective C: Finally, we will make our tools available to other university teams by deploying a web platform. This platform must be enough scalable to process a large number of data. Moreover, security and privacy of the data must be ensured. Tools developed for the previous objectives will be integrated to the platform.