Motion Integrated AI-based Cardiac MRI Reconstruction – MEDICARE

Accelerated MRI reconstruction techniques are necessary to fasten long cardiac exams and avoid the use of breath holds. Parallel imaging (PI) techniques are widely used in clinical practice to reduce the number of k-space lines acquired, by taking advantage of multiple receiver coil information. GRAPPA is a PI reconstruction algorithm aiming at interpolating the missing k-space lines using convolution kernels determined during a calibration based on fully sampled central k-space lines. RAKI algorithm recently proposed to incorporate non-linearity by applying CNN with multiple layers. The advantage of this approach lies in the single shot training, which is uncommon for DL techniques usually requiring large training dataset.

We first assessed the need of multiple CNN layers and non-linear activation for a RAKI reconstruction. As such, we propose the use of a single layer RAKI network, that comes close to a parallelised GRAPPA algorithm to further accelerate the reconstruction, while guaranteeing imaging quality.
In a second step we proposed to accelerate the training of the CNN weights and using the redundancy of information between the cardiac phases and adjacent slices.

The PSNR of the proposed accelerated method is almost as good as Standard RAKI and GRAPPA, while the reconstruction time is reduced.

The reconstruction on three adjacent slices showed the best compromise between image quality and reconstruction acceleration.

GRAPPA, Standard RAKI and our method seem to show comparable visual results

Artefacts remain for several examples during evaluation of the RAKI-based approaches. These artefacts seem to be present on standard RAKI and intrinsic of the technique. Future works will pinpoint these artefacts in the hope of reducing or eliminating them. One assumes that these artefacts seem to be due to overtraining on the calibration data, and structures of the subsampling. One will investigate the use of more raw-data specific loss functions, and/or the use of non-cartesian subsampling.

.

Magnetic Resonance Imaging (MRI) is the recommended medical imaging modality for a wide range of cardiovascular pathologies, and the use of cardiac MRI in clinical practice is only going to be even more widespread given the pace of technology developments in the last years.
Deep Learning (DL) and other types of so-called Intelligence Artificial (IA) techniques have recently hit the computer vision communities and has allowed reaching levels of performance on par with humans for several tasks (object classification, … ). These techniques have therefore naturally been translated to the medical imaging field with similar successes, although specific limitations are inherent to the medical fields (such as the relatively scarceness of annotated datasets). More recently, DL-based medical imaging reconstruction techniques have been proposed, as they could allow for a faster acquisition and reconstruction data, and therefore speedup MRI examination and lessen their costs on healthcare.
The MEDICARE project aims at bringing the expertise of both partners, MRI reconstruction and physiological data analysis for the IADI lab and Medical Imaging analysis for the IMI, in order to develop a Motion Integrated AI-based cardiac MRI reconstruction technique allowing for whole heart cardiac imaging during free-breathing. Such a solution would open up a vast range of clinical applications (accelerated high resolution CINE images, accelerated acquisition of T1/T2 maps). Overcome certain limitations need to be overcome before transferring such AI-based solution into clinical practice, mostly based on the need to gain radiologists trust by building robust solutions (avoiding oversimplified regularization biased towards healthy tissues) and providing confidence maps along with reconstructed images to help radiologist taking a rational diagnosis decision.
The overall goal of developing a Motion-Integrated AI-based cardiac reconstruction technique will be achieved through the following four sub-objectives:
• Develop accelerated AI-based MRI reconstruction that incorporate uncertainty modelling into the analysis to enhance the trustworthiness of image quality
• Advance MRI reconstruction of subtle abnormal anatomical details to avoid bias of healthy tissues in AI-models and improve clinical reliability
• Develop and incorporate AI-based 4D motion prediction without full reconstruction for faster sparse k-space sampling for MRI acquisition
• Enhance the explainability of AI methods for MR image processing by combining AI based feature extraction with robust graph-based optimization strategies