Introduction

Prostate diffusion-weighted (DW) MRI are used as part of multi-parametric (mp) MRI for detecting tumours. DW-MRI, based on single-shot echo planar imaging (EPI), can suffer from geometric distortions caused by off-resonant magnetic fields arising from susceptibility differences between prostate tissue and rectal air. Previously, Usman et al^1,2 introduced two model-based reconstruction approaches for correcting distortions, including signal pile-ups, and were demonstrated on low resolution DW data with b-values of 0 and 500s/mm­² only. The correction process requires two opposite phase-encoded DW-EPI datasets with equal number of measurements and a B₀-field map. The aim of this work is to demonstrate the application of one of the model-based reconstruction approaches¹ on typical clinical prostate DW-MR datasets.

Methods

Image acquisition: Two prostate MR patients (60-70y.o.) consented for additional acquisitions. Four MR scans (T2-weighted (T2W), B₀-field and two DW scans) were acquired within 17minutes. (1)An axial T2W image with turbo spin echo, resolution=0.6×0.6×3mm³, field of view (FOV)=220×220×57mm³, SENSE=1.3, TE/TR=100msec/4000msec. (2)A B₀-field map was calculated using a 3D dual-echo gradient-echo scan with resolution=1.3×1.3×1.3mm³, FOV=230×230×92mm³, partial-Fourier acquisition=0.78 (slice direction), TE/TR=4msec/9msec. (3)A typical clinical prostate DW-MR protocol with single-shot EPI, resolution=1.3×1.3×5mm³, FOV=220×220×55mm³, partial-Fourier acquisition=0.63, SENSE=1.6, phase‐encode bandwidth/pixel=42.6Hz/pixel, TE/TR=80msec/2600msec, b-values of 0s/mm² (Number of signal averages (NSA)=6), 150s/mm² (NSA=6), 500s/mm² (NSA=12) and 1000s/mm² (NSA=18) each with three orthogonal diffusion directions and phase‐encoding direction in the anterior‐posterior (AP) direction (DW-P). Only half of the NSAs from this DW-P dataset were used for model-based reconstruction. (4)The same DW sequence but with opposite phase-encoding direction (DW-A) and half of the NSAs of DW-P was also acquired for model-based reconstruction. All sequences used first order volume shimming on the prostate region and its vicinity.
Data post-processing: Data was processed similarly to¹ and is illustrated in Fig.1. However, some modifications were made to the pipeline. The large raw EPI datasets (approx.11GB per patient) were reduced 4x in size using array coil compression³ to enable preprocessing on a MacbookPro (16GB RAM). Information about the shim gradients was extracted from the DICOM files and the B₀-field maps were modified to account for differences in shim gradient induced B₀-fields between DW and B₀ scans.
To compensate for frequency changes between the DW-P and DW-A scans, a reliable fixed center frequency offset was computed using an iterative optimisation process with modified objective function and multiple start points (ranging from -40 to 40Hz). Phase correction for the DW datasets with b-values >0s/mm² were carried out to minimise SNR loss during averaging. Due to the high-resolution data with many b-values and NSAs, model-based reconstruction was performed using University College London’s research computing platform, where correction of each individual measurement (DW-P and DW-A pair) using 16GB RAM and 12 parallel processors took approx.15 minutes.
Image analysis: Uncorrected ADC maps (using DW-P NSA=6) and corrected ADC maps (using DW-P NSA=3 and DW-A NSA=3) were calculated for all slices of the prostates and qualitatively compared to the T2W images resampled to the same resolution and slice thickness. Prostate ROIs were drawn on the T2W images, uncorrected and corrected ADC datasets. The Dice similarity scores with respect to the T2W images were calculated. The uncorrected and corrected isotropic b=1000s/mm² were compared visually with the T2W image and the signal to noise ratio (SNR) at the prostate was calculated using the difference method⁵.

Results

Fig.2 demonstrates that the corrected ADC maps have higher Dice scores and prostatic zone similarity to T2W images than the uncorrected ADC maps. Fig.3 shows that corrected isotropic b=1000s/mm² have higher SNR (7.2±1.3 and 3.0±0.3) than uncorrected data (3.9±0.5 and 2.1±0.4) suggesting the phase correction step is effective. However, some blurring is observed in the corrected b=1000s/mm² data. A further comparison (Fig.4) of all b-values for a different corrected slice of Patient 1 shows visibly increasing shifts in the prostate in the phase-encoding direction with increasing b-values, which suggests scanner frequency drifts and potential contribution from eddy currents.

Discussion

Early data indicates that distortion correction using model-based reconstruction for clinical standard dataset is challenging but feasible. The benefits are distortion correction including pile-up correction, improved DICE score (useful for automatic processing of mp-MRI) and significant increases in SNR of high b-value DW data - all for an extra scan time of 3minutes (the B0 scan). However, clinical data are usually large leading to lengthy postprocessing times. Another limitation is the static center frequency offset computed using the first b=0s/mm² from DW-P and DW-A. Further frequency drifts are not currently considered for b-values that are acquired up to 4minutes after the initial b=0s/mm² and may explain the blurring observed in Fig.4. Possible additions of extra b=0s/mm² measurements at intervals, accounting for eddy currents and image acquisition using interleaved double spin-echo sequence⁶ may provide more accurate B₀-field maps to prevent blurring. Pile-up correction is not always effective and future work aims to investigate the gradient of the B₀ map, the B₀ map smoothing applied and methods such as² for refining the B₀ map.

Conclusion

Distortion correction of prostate DW-MRI with typical clinical scan parameters using model-based reconstruction is feasible but challenging. Potential one-to-one mapping of ADC maps to T2W images may improve mp-MRI but tighter control of B₀ fluctuations and possibly eddy current effects may be required.

Acknowledgements

This work is supported by Cancer Research UK/EPSRC (grant A21099) and the NIHR Biomedical Research Unit at UCH. We acknowledge research support provided by Philips Healthcare. The authors would also like to thank Elizabeth Isaac for recruiting the patients.