AI-assisted real-time motion correction in magnetic resonance
|Tommaso Torda||Dottorando/Assegnista/Specializzando componente non strutturato del gruppo di ricerca|
|Riccardo Faccini||Componenti strutturati del gruppo di ricerca|
|Carlo Mancini Terracciano||Componenti strutturati del gruppo di ricerca|
|Cecilia Voena||Ricercatrice||Istituto Nazionale Fisica Nucleare||Altro personale aggregato Sapienza o esterni, titolari di borse di studio di ricerca|
Magnetic Resonance Imaging (MRI) is an image modality highly susceptible to performance degradation due to subject motion. Artefacts caused by patient movements are a critical problem in many clinical and research MRI applications (involuntary subject's movement, no-collaborative patients affected by disease with tremor, fetal MRI). The result is a loss of image resolution, which prevents early diagnosis and increases the costs and time of healthcare due to repeated examinations.
This project proposes a novel high speed real-time prospective motion correction technique based on modern Deep Learning (DL) methods which provide a potential avenue for dramatically reducing the computation time and improving the convergence of retrospective motion correction overcoming the limits of the current state-of-the-art prospective methods.
Exploiting cutting-edge developments from the field of experimental high energy physics that allow to run sophisticated algorithms for reconstruction and inference of complex data at speeds much faster than the typical acquisition time of MRI sequences, and the know-how in developing and training innovative deep neural network (DNN) models in both foundamental and applied physics in medical imaging, the technique we propose has the potential to outclass current state-of-the-art motion correction methods with ground-breaking advancement in the clinical diagnostic based on MRI.
Innovative DNNs for motion correction in MRI will be developed using publicly available datasets. To fulfil the tight requirements of high accuracy and low latency required by the motion correction task, we will implement DNN compressione and simplification methods developed in high energy physics. The models will be then deployed into real-time co-processors, such as Field Programmable Gate Arrays (FPGAs) and GPUs, and performances with respect the motion-correction task will be evaluated.