Cardiac MR function analysis with DL-based super resolution reconstruction: application in the clinical setting

Abstract To assess differences in volumetry, image quality and acquisition time between balanced steady-state free precession cine sequences acquired using (a) a standardized sensitivity encoding (SENSE) approach and (b) deep learning-based super-resolution reconstruction based on high-resolution images acquired with compressed sensitivity encoding (C-SENSE). We retrospectively evaluated 31 consecutive patients (mean age 61.2 ± 13.1 years, 26% female (8/31) and 74% male (23/31)) undergoing cardiac magnetic resonance imaging (MRI) examinations to assess for the presence of ischemic and non-ischemic cardiomyopathies. Cine images were acquired using a 1.5T Philips Ingenia MRI scanner, with classic parallel imaging (SENSE) and compressed sensing (C-SENSE) accelerated acquisition techniques ( R = 2 and R = 4, respectively). C-SENSE datasets were reconstructed using a deep learning-based denoising and super-resolution algorithm to enhance image resolution and quality (CS-SR). To evaluate cardiac function, manual left ventricular (LV) segmentation and volumetric analysis were performed on both datasets by two readers, who were blinded to the clinical data. Image quality was rated independently by three readers using Likert scales. Correlation between SENSE and CS-SR datasets with respect to LV volumetry was high ( r = 0.98-1.00), with no significant differences found for end-diastolic volume (mean difference 0.04 ml, limits of agreement (LoA) -11.19 to 11.26 ml; p = 0.970) or end-systolic volume (mean difference 1.60 ml, LoA − 7.48 to 10.68 ml; p = 0.064). Overall subjective image quality was comparable ( p = 0.061), with CS-SR offering better image sharpness at the cost of increased artifacts ( p < 0.001 respectively). Image acquisition time was significantly accelerated with C-SENSE acquisition (SENSE: 411.1 ± 47.7 s, C-SENSE: 165.6 ± 21.5 s; p < 0.001). CS-SR shows promise in streamlining routine cardiac imaging by significantly shortening acquisition times, without impairing LV volumetric analysis, while preserving overall image quality and resolution.