OP02.03: AI supported super‐resolution image reconstruction in fetal MRI

Any technology improving image quality allows for a more accurate and reliable clinical integration of fetal MRI as a diagnostic tool in prenatal medicine. The clinical utility of novel k-space reconstruction techniques are currently under investigation in different fields of radiology, however have not yet been assessed in fetal MRI. This ongoing prospective study compares compressed sensing and deep learning/convolutional neural network-based reconstruction methods in fetal MRI. Conventional T2-TSE sequences were acquired in a series of 30 clinically indicated fetal MR examinations at 3 Tesla, including cases with normal brain development and cerebral malformations. The same acquired data underwent three reconstruction methods: 1) compressed sensing reconstruction incorporating parallel imaging coil information (CS) [Meister 2022]; 2) sparsity constrained reconstruction [Pezzotti 2020] combining CS and deep learning (CSDL); 3) additional convolutional neural network - CNN (Precise-Image-Net) [Kim 2015] combining CSDL for removal of ringing artifacts and upscaling of image resolution (CSDL2). Different reconstruction methods were compared and rated according to detail/resolution, overall image quality and contrast. AI supported image reconstruction techniques were successful in several examinations. Initial experiences in 30 fetuses (study ongoing) indicate a higher detail and decreased noise after using CS and deep learning as well as CNN-based k-space reconstructions. Upscaling of image resolution lead to higher contrasts between arteries in the subarachnoid space and surrounding CSF as well as better delineation of the fetal vermis, optic nerves and olfactory bulbs. Initial experiences indicate an improvement of diagnostic detail and image quality after using novel k-space derived image reconstructions.