1351999 - OpenHeart - Open-source raw data repository for cardiac MRI

Machine learning has strongly improved MR image reconstruction in recent years. Especially for cardiac cine MRI, a wide range of techniques have been proposed yielding excellent image quality despite high undersampling factors. These methods are commonly supervised techniques which require a high number of datasets. Open-access databases for e.g. knee or brain MR raw data have been presented [1, 2]. For cardiac applications a large open-access, community-led database of MR raw data is still needed. Existing cardiac MR raw data (e.g. [3]) need to be pooled from different institutions to establish a diverse database populated by the entire cardiac MR community. This will ensure a fair and unbiased distribution of data to facilitate an objective comparison and boost future developments.

Methods:

Here we present a framework for an open-access database where users from different institutions can upload cardiac MR raw data. The framework consists of three main components (Fig. 1): (i) a database based on XNAT [4] which allows users to view and download datasets, (ii) a web-based interface which enables users to upload cardiac MR raw data in MRD (ISMRMRD) format and (iii) a quality control (QC) module which verifies the integrity and quality of the uploaded raw data. The QC module carries out a dockerized image reconstruction and detects any technical errors in the data, e.g. missing reference lines for accelerated parallel imaging acquisition. Reconstruction is carried out in a docker container using the Synergistic Image Reconstruction Framework (SIRF) and Gadgetron [5]. The reconstructed images are presented to the user for visual inspection. The data are stored in the database only after the user confirms the image quality. Currently cardiac cine MRI with a Cartesian GRAPPA sampling pattern is supported. Further acquisition schemes such as non-Cartesian sampling patterns (e.g. radial and spiral) are under development.

Results:

The reconstructed images presented to the user to verify the image quality are shown in Fig. 2. Subsequently, the user can select which scans should be stored in the database. An example of a dataset is shown in Fig. 3. The scan parameters (i.e. meta data) such as image resolution are automatically extracted from the MRD file for each dataset. This allows users to search through parameters stored in the raw MRD file and select e.g. only scans with a certain voxel size. The reconstructed QC images are also available to the user for reference purposes.

Conclusion:

We have developed a framework which allows the cardiac MR community to collate raw data to boost future development of image reconstruction algorithms, especially for machine-learning-based algorithms. The aim is to also have part of the data not open-access to facilitate independent evaluation (e.g. in the form of challenges). This database can also open the field of image reconstruction to scientists who currently do not have access to an MR scanner allowing for more inclusive research.