Achieving Ultra-High Acceleration Rates in 7T MRI Using Combined Controlled Aliasing in Parallel Imaging and Compressed Sensing with Deep-Learning-Based Image Reconstruction.
Authors
Affiliations (2)
Affiliations (2)
- From the Department of Radiology (E.H.M., X.Z., S.T., V.N.P., J.O.E., E.M.W., V.G.), Neurologic Surgery (E.H.M.), Mayo Clinic, Jacksonville, Florida; Swiss Innovation Hub (T.Y., G.F.P.), Siemens Healthineers International AG, Lausanne, Switzerland; Department of Radiology (T.Y.), Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland; LTS5, Ecole Polytechnique Fédérale de Lausanne (T.Y.), Lausanne, Switzerland and Research & Clinical Translation (D.N.), Magnetic Resonance, Advanced Systems (J.H., P.L.), Magnetic Resonance, Siemens Healthineers AG, Erlangen, Germany. [email protected].
- From the Department of Radiology (E.H.M., X.Z., S.T., V.N.P., J.O.E., E.M.W., V.G.), Neurologic Surgery (E.H.M.), Mayo Clinic, Jacksonville, Florida; Swiss Innovation Hub (T.Y., G.F.P.), Siemens Healthineers International AG, Lausanne, Switzerland; Department of Radiology (T.Y.), Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland; LTS5, Ecole Polytechnique Fédérale de Lausanne (T.Y.), Lausanne, Switzerland and Research & Clinical Translation (D.N.), Magnetic Resonance, Advanced Systems (J.H., P.L.), Magnetic Resonance, Siemens Healthineers AG, Erlangen, Germany.
Abstract
Clinical adoption of 7T MRI has been limited by lengthy acquisitions. Acceleration techniques, such as controlled aliasing in parallel imaging (CAIPI) and compressed sensing (CS), can reduce scan time but are prone to artifacts and noise. This work combines CAIPI and CS within a unified, two-step deep-learning (DL) reconstruction framework to leverage their strengths: CAIPI for controlled aliasing with improved conditioning, CS for additional incoherent undersampling, and DL to reduce residual aliasing and CS-related artifacts. This hybrid approach aims to maintain image quality while enabling higher net acceleration and reducing scan time. In this paired within-subject study, 30 patients underwent 7T sampling perfection with application-optimized contrasts using different flip angle evolutions (SPACE) FLAIR and 30 patients underwent 7T SPACE T2 acquisitions. Each scan included a reference CAIPI-DL protocol (acceleration factor=6; ∼7 minutes) and a CS-CAIPI-DL protocol (acceleration factor=14; ∼3.5 minutes). Image quality was assessed quantitatively using structural similarity index, peak signal-to-noise ratio, gradient-domain mean squared error, contrast-to-noise ratio, noise, ghosting ratio, and Natural Image Quality Evaluator (NIQE), as well as blinded qualitative assessments. Statistical comparison used linear mixed-effects models with false discovery rate correction. CS-CAIPI-DL reduced scan time by approximately 50% without significant differences in contrast-to-noise ratio or noise for either sequence. Ghosting ratios were significantly lower with CS-CAIPI-DL for both FLAIR (∼10% reduction; q<.001) and T2 (∼30% reduction; q<.001), consistent with reduced motion-related artifact. NIQE scores were significantly improved for T2 (q<.001) and showed a favorable trend for FLAIR (q=.06). There was no significant difference in diagnostic adequacy and no DL-specific artifacts. Combining CS with CAIPI and DL reconstruction enables over two-fold increase in acceleration with no statistically significant differences in key image quality metrics. Shorter acquisitions were associated with significantly lower ghosting ratios, consistent with reduced motion-related artifact, supporting the feasibility of rapid, high-resolution 7T brain MRI.