Distortion-Free High-Resolution PROPELLER-DWI: Technical Advances and Initial Demonstration for the Prostate

Introduction

• Off-resonance during the long single-shot EPI readout causes geometric distortion, signal pile-up, and signal loss, especially in the prostate.Rectal gas, bowel contents, postoperative changes, and metallic implants can produce severe warping.Source: Introduction · p. 3
• High-resolution EPI-DWI is technically challenging because smaller voxels reduce SNR while longer readouts increase off-resonance distortion and prolong TE.The larger readout matrix lengthens echo spacing within gradient hardware limits.Source: Introduction · p. 3
• Existing EPI-based strategies can improve geometric fidelity, but residual off-resonance distortion may persist in challenging pelvic imaging and multi-shot methods can increase scan time.Time-resolved EPI approaches require model-based spatiotemporal reconstruction, adding complexity and burden.Source: Introduction · pp. 3–4
• The proposed PROPELLER-DWI framework combines PGSE diffusion encoding, phase-insensitive GRASE PROPELLER acquisition, and self-supervised joint reconstruction and denoising.The authors evaluate it against clinical single-shot EPI-DWI across spatial resolutions, including geometric fidelity, image quality, radiologist assessment, ADC quantification, and retrospective reduced-shot reconstruction.Source: Introduction · p. 4

PROPELLER-DWI Acquisition Design

• Each shot acquires a GRASE train following a PROPELLER sampling pattern, with blade width set by spin-echo train length and acceleration.The number of gradient echoes determines how many blades are acquired per shot.Source: PROPELLER-DWI Acquisition Design · p. 5
• PGSE diffusion encoding precedes the readout, and a phase-insensitive preparation is applied because physiological motion can disrupt the CPMG condition.After PGSE, a dephasing gradient and 90° pulse rotate the non-CPMG component out of the transverse plane and suppress incoherent signal contributions.Source: PROPELLER-DWI Acquisition Design · p. 5
• Matched rephasing gradients are applied during the subsequent refocusing echo train.This follows the phase-insensitive preparation described for the diffusion-encoded signal.Source: PROPELLER-DWI Acquisition Design · p. 5
• Successive shots use distinct blade orientations, rotated by a golden-angle increment to build complementary k-space coverage.For N uniformly spaced blades, the pattern's rotational symmetry interval is 180°/N, and the rotation increment is that interval divided by the golden ratio.Source: PROPELLER-DWI Acquisition Design · p. 5

Self-Supervised Joint Reconstruction and Denoising

• The self-supervised scheme trains an unrolled network to reconstruct and denoise PROPELLER-DWI from undersampled multicoil k-space data.Acquired data are split into complementary subsets, with one subset as network input and the other used for self-supervised data-consistency loss.Source: Self-Supervised Joint Reconstruction… · pp. 5–6
• Blade-wise phase correction is performed before k-space splitting by removing the low-pass phase from each full-resolution blade image.The phase is estimated from a Gaussian low-pass-filtered blade image.Source: Self-Supervised Joint Reconstruction… · p. 6
• The network predicts data consistent with held-out measurements, and its parameters are optimized using a mixed image-domain loss.The loss compares the network reconstruction projected onto the held-out sampling operator with the image reconstructed from held-out data.Source: Self-Supervised Joint Reconstruction… · p. 6
• The reconstruction used 12 cascaded iteration blocks alternating data-consistency updates and learned regularization; coil-sensitivity maps were estimated from acquired low-resolution calibration k-space using a 4-layer U-Net and jointly optimized during b=50 s/mm2 image training.Source: Self-Supervised Joint Reconstruction… · p. 6

Phantom Imaging Experiments

• The NIST/QIBA phantom contains 13 cylindrical vials, including water and aqueous PVP solutions at five concentrations.The PVP concentrations are 10%, 20%, 30%, 40%, and 50%.Source: Phantom Imaging Experiments · p. 7
• The PVP solutions provide characterized isotropic ADC values for assessing diffusion-measurement accuracy, bias, and linearity.The phantom was allowed to reach thermal equilibrium for approximately 2 hours before imaging, at approximately 19 ºC.Source: Phantom Imaging Experiments · p. 7
• Phantom imaging compared EPI-DWI at 1 × 1 mm2 and 2 × 2 mm2 in-plane resolutions with PROPELLER-DWI at 1 × 1 mm2.All acquisitions used a 3 mm slice thickness.Source: Phantom Imaging Experiments · p. 7

In Vivo Imaging Experiments

• The study included 25 male volunteers and 4 male patients, with patient scans treated as feasibility examples and cross-validation, reader assessment, and statistical comparisons performed in volunteers.The volunteer cohort had a mean age of 41.9 ± 16.9 years and the patients 69.2 ± 9.1 years.Source: In Vivo Imaging Experiments · p. 7
• Volunteers underwent EPI-DWI and prototype PROPELLER-DWI after axial T2-weighted imaging, using three resolution settings and diffusion-weighted acquisitions at b=50 and b=1000 s/mm2.Three orthogonal diffusion directions were acquired at each b-value.Source: In Vivo Imaging Experiments · pp. 7–8
• In-vivo PROPELLER-DWI acquisitions with different gradient echoes were evaluated while total blades were fixed at 36; the six-gradient-echo configuration was selected for subsequent experiments to balance scan time and off-resonance artifacts.Source: In Vivo Imaging Experiments · p. 8
• The PROPELLER-DWI protocol used 5 shots at b=50 s/mm2 and 18 shots at b=1000 s/mm2, for a total acquisition time of 6 min 50 s.The evaluated spatial resolutions were 1.8 x 1.8 mm2, 1.4 x 1.4 mm2, and 1.25 x 1.25 mm2, compared with clinical EPI-DWI at 2.0 x 2.0 mm2.Source: In Vivo Imaging Experiments · p. 8
• Only the 1.25 x 1.25 mm2 PROPELLER-DWI protocol was acquired in the four patients to minimize disruption to their clinical MRI exams.Two patient reports identified PI-RADS 3 lesions.Source: In Vivo Imaging Experiments · pp. 7–8

Conventional Reconstruction and MPPCA Denoising

• Missing phase-encoding lines within each blade were recovered using GRAPPA, after which blades with matching b-values and diffusion directions were combined using NUFFT.The reconstruction estimated and removed low-frequency image phase attributed to physiological motion and gradient-echo dephasing before combining blades.Source: Conventional Reconstruction and MPPCA… · p. 8
• Density-compensation weights were calculated separately for each diffusion-encoding direction from its sampling trajectory using adjoint NUFFT.Unit values were assigned at sampled non-Cartesian k-space locations to obtain the image-space weighting, and the forward NUFFT of its reciprocal was used as the weights.Source: Conventional Reconstruction and MPPCA… · p. 8
• Diffusion trace images were generated from the three-direction images, and ADC maps were calculated from the b=50 and b=1000 s/mm2 images.Source: Conventional Reconstruction and MPPCA… · p. 9
• MPPCA denoising was applied to raw data along the coil dimension for each blade before the denoised data underwent the same GRAPPA and NUFFT pipeline.Source: Conventional Reconstruction and MPPCA… · p. 9

Self-Supervised Training

• Five repeated evaluations used train-test rounds across 25 in vivo datasets, holding out five unique subjects and one common representative subject for inference in each round.An independent model was trained on the remaining 19 subjects, so each held-out subject was reconstructed with a model not trained on that subject.Source: Self-Supervised Training · p. 9
• During training, the k-space split ratio was randomly varied from 0.3-0.99 to expose the network to a broad range of undersampling levels.Source: Self-Supervised Training · p. 9
• Inference applied the trained model to the full acquired k-space data without splitting.Source: Self-Supervised Training · p. 9
• Retrospective reduced-shot reconstructions used 3/12 and 2/6 shots for b=50/b=1000 s/mm2, corresponding to total acquisition times of 4 min 30 s and 2 min 29 s, respectively.Full-shot reconstructions used 5 shots at b=50 s/mm2 and 18 shots at b=1000 s/mm2.Source: Self-Supervised Training · p. 9

Image Quality Assessment

• Two fellowship-trained body radiologists independently evaluated EPI-DWI and PROPELLER-DWI images displayed side-by-side in randomized left-right order, scoring overall quality, sharpness, and artifacts and recording a preference for each paired comparison.Scores used a 5-point Likert scale, with higher scores indicating better image quality.Source: Image Quality Assessment · pp. 9–10

Geometric Fidelity Assessment

• Geometric fidelity was rated separately using b=50 s/mm2 DWI paired with the corresponding T2-weighted image as the anatomic reference.The 4-point ordinal scale ranged from severe distortion to no visible distortion, shape change, or signal pile-up.Source: Geometric Fidelity Assessment · p. 10
• Scan-to-scan motion and nonrigid anatomic changes, including bladder filling and bowel-related displacement, limited reliable registration- or segmentation-based distortion quantification.Source: Geometric Fidelity Assessment · p. 10

Statistical Analysis

• Paired Wilcoxon signed-rank tests compared image-quality scores by resolution and image type, and geometric-fidelity scores by resolution; Benjamini-Hochberg correction was applied separately.Image types were b=50 s/mm2, b=1000 s/mm2, and ADC.Source: Statistical Analysis · p. 10
• Reader preference was evaluated within bins of average EPI-DWI geometric-fidelity scores using exact binomial tests.The preferred method was defined by both readers within each bin.Source: Statistical Analysis · p. 10
• ADC resolution dependence was tested with subject-specific random-intercept mixed models and likelihood-ratio tests, with Holm correction across four comparisons.The analysis used four manually placed ROIs per subject, covering left and right peripheral and transition zones.Source: Statistical Analysis · pp. 10–11
• Inter-reader agreement was reported using weighted kappa, while left-right ADC differences were assessed with paired t-tests.Source: Statistical Analysis · pp. 10–11

Results

• MPPCA denoising reduced noise visibly at low resolution but was less effective at higher spatial resolution, where SNR was lower and fine detail more vulnerable to noise.Source: Results · p. 11
• Across resolutions, model-to-model SSIM remained 0.967-0.985 for b=50 s/mm2 and 0.948-0.976 for b=1000 s/mm2.Difference maps showed model-dependent variation.Source: Results · p. 11
• With rectal gas, EPI-DWI distortion and signal degradation worsened particularly at higher resolution, while PROPELLER-DWI preserved prostate geometry and interpretable images across resolutions.Source: Results · pp. 11–12
• With metallic fiducial markers, EPI-DWI artifact worsened at high b-value and high resolution, while PROPELLER-DWI reduced artifact and preserved prostate anatomy.Source: Results · p. 12
• PROPELLER-DWI ADC measurements were consistent across resolutions, whereas in-vivo EPI-DWI showed significant resolution-dependent ADC trends.In the phantom, PROPELLER-DWI ADC agreed with ground-truth and EPI-DWI values.Source: Results · p. 12
• PROPELLER-DWI scored higher than EPI-DWI for image quality, sharpness, and artifact level; artifact scores were significantly higher across all image types and resolutions.Overall image quality and sharpness were significantly higher in most comparisons, with exceptions at low-resolution b=1000 s/mm2 and highest-resolution ADC.Source: Results · p. 13
• PROPELLER-DWI had significantly higher geometric-fidelity scores than EPI-DWI across all three spatial-resolution settings.The difference indicated less visible distortion and prostate shape deformation.Source: Results · p. 13
• At 1.25 x 1.25 mm2 resolution, reduced-shot reconstructions remained visually similar to the full-shot reference, with increasing noise as shots decreased.The 4 min 30 s and 2 min 29 s reconstructions showed small ADC differences; the shorter reconstruction had greater noise and larger difference-map changes but preserved overall prostate anatomy and ADC.Source: Results · p. 13

• PROPELLER-DWI preserved prostate geometry where single-shot EPI-DWI showed susceptibility distortion, signal loss, or metal artifacts, while supporting higher resolution and consistent ADC across resolutions.Source: Discussion · p. 14
• Repeated spin-echo refocusing reduces off-resonance geometric distortion, while GRASE improves acquisition efficiency and lowers SAR versus FSE-only approaches, with residual gradient-echo off-resonance sensitivity.Source: Discussion · p. 14
• PROPELLER-DWI maintained consistent TE across resolutions, avoiding EPI's added high-resolution TE penalty and enabling clinically feasible high-resolution imaging despite voxel-size SNR loss.Source: Discussion · pp. 15–16
• Self-supervised joint reconstruction and denoising reduced noise while preserving fine anatomy and outperformed MPPCA, especially at higher spatial resolution.Source: Discussion · p. 16
• Readers more often preferred PROPELLER-DWI when EPI distortion was substantial, whereas preference was more balanced with mild distortion; it may complement high-resolution or technically challenging exams.Source: Discussion · p. 16
• EPI-DWI ADC decreased with increasing spatial resolution, particularly in the transition zone, while PROPELLER-DWI ADC remained consistent across resolutions but was slightly higher than EPI-DWI.Source: Discussion · pp. 16–17

Discussion

• Slightly higher in vivo ADC with PROPELLER-DWI may reflect motion-induced signal attenuation that mimics additional diffusion and overestimates ADC.The passage contrasts this observation with low-resolution EPI-DWI.Source: Discussion · p. 17
• SNR remains limited at high b-values and spatial resolutions; self-supervised reconstruction mitigated this, but larger, more diverse training data may help further.The phase-insensitive preparation discards approximately half the signal; future work will investigate optimized phase cycling and non-CPMG signal stabilization to recover signal pathways and improve SNR efficiency27,28,52.Source: Discussion · p. 17
• Fully sampled PROPELLER-DWI takes longer than single-shot EPI-DWI, although retrospective reduced-shot reconstruction demonstrated feasibility of substantially shorter scans.The study did not systematically compare advanced EPI approaches, including reduced-FOV or multi-shot EPI, which can reduce distortion and achieve shorter TE than the evaluated high-resolution single-shot EPI protocols.Source: Discussion · p. 17
• Centric echo ordering causes symmetric T2-dependent low-pass weighting and blurring within each blade, but overlapping rotated-blade contributions determine the final image point-spread function.Future work will investigate linear ordering and model-based T2-blur correction during reconstruction.Source: Discussion · pp. 17–18
• The gradient-and-spin-echo readout interleaves gradient-echo encoding with refocusing RF pulses and may reduce readout gradient duty cycle, warranting study of nerve stimulation and patient comfort.These practical implications need investigation in broader clinical implementation54.Source: Discussion · p. 18
• The reader study assessed geometric fidelity, image quality, and preference, so larger prospective studies with PI-RADS-based assessment and histopathologic correlation are needed to test diagnostic benefits.The needed outcomes include lesion detection, localization, and risk stratification.Source: Discussion · p. 18

Conclusion

• The framework combines golden-angle rotating interleaved GRASE PROPELLER acquisition with self-supervised joint reconstruction and denoising.Source: Conclusion · p. 18
• The framework enables higher spatial resolution while preserving geometric fidelity within a clinically feasible scan time compared with the current clinical standard.Source: Conclusion · p. 18
• Retrospective reduced-shot reconstruction demonstrated the potential for further scan-time reduction.Source: Conclusion · p. 18
• Radiologists preferred PROPELLER-DWI particularly in cases where EPI-DWI suffered from geometric distortion.Source: Conclusion · p. 18
• The authors support PROPELLER-DWI as a potential alternative for high-resolution quantitative and diagnostic prostate DWI, particularly in technically challenging pelvic imaging settings.Source: Conclusion · p. 18