Introduction

Signal averaging is normally used in diffusion-weighted imaging (DWI) due to low SNR. Compared to magnitude signal averaging, complex signal averaging can reduce the noise floor of DWI images, therefore improving contrast and the estimation of diffusion parameters [1-3]. However, the robustness of complex signal averaging highly depends on the performance of phase correction for the removal of shot-to-shot background phase variations. Low-pass filters are commonly used for phase estimation assuming the background phase is smooth. However, motion, field inhomogeneity, etc. could create rapid phase changes. After applying low-pass filters, these rapid phase changes are lost, causing signal cancellation (wormhole artifacts) in the complex averaged images. Reducing filter kernel size could minimize wormhole artifacts, but it also undermines the efficiency of noise reduction. Recently, we developed a deep-learning-based phase correction (DLPC) method for robust complex averaging in DWI [4]. In this work, we combined DLPC with deep learning reconstruction for robust real diffusion-weighted MRI to improve diffusion contrast in the acquired and synthesized high-bValue images.

Methods

To evaluate the noise floor, a double-wall breast phantom (CALIBER MRI) was scanned using 10 b-values, 50, 200,400, 600, 800, 1000, 1250, 1500, 2000, 2500 s/mm2 with a NEX of 1, 1, 1, 1, 1, 1, 2, 4, 5, 6, imaging matrix, 120x120, and echo time, 67.7ms. Volunteer scans were performed on various 1.5T and 3.0T scanners (GE Healthcare) with IRB approval and written informed consent. Neural DWI was performed on a 3.0T scanner (750w) with a 24-ch head coil, b=5000s/mm2; NEX=10; imaging matrix=160x160; echo time=114.2ms. The routine prostate DWI of a healthy volunteer was performed on a 1.5T scanner (Artist) with 30-ch AIR coil; b=50 (NEX=2) and 800 (NEX=12) s/mm2; imaging matrix=80x40; echo time=66.5ms. Synthetic b-values of 1400 and 2000s/mm2 were generated from the acquired b-values. High b-value (2000s/mm2) prostate DWI of two healthy volunteers were performed on a 3T scanner (Architect) with 21-ch AIR coil; NEX=15; imaging matrix=96x96; echo time=91.6ms, and on a 1.5T scanner (Voyager) with 30-ch AA AIR coil; NEX=20; imaging matrix=74x38; echo time=84.5ms. An accelerated prostate DWI of a healthy volunteer was performed on a 1.5T scanner (Voyager) with 16-ch coil; b=50 (NEX=2), 800 (NEX=7) and 1400 (NEX=14) s/mm2; imaging matrix=90x90; echo time=82.9ms; acceleration factor=2. DL-based phase correction [4] and deep learning reconstruction [5] were embedded in the reconstruction pipeline to generate DL PC DWI images from the originally acquired raw MR data.

Results and Discussion

The phantom solutions are known to have mono-exponential diffusion. As shown in Fig.1b, the DWI signal exhibits mono-exponential decay in a slow-diffusing solution due to sufficient SNR. However, the DWI signals are close to the rectified noise floor and deviate from the mono-exponential decay in moderate- and high-diffusing solutions (Fig.1b, c) at high b-values. With DLPC, DWI signal decay (Fig.1, purple curves) is back to monoexponential decay due to the efficient reduction of the noise floor. Besides a high noise floor (Fig2.a), the wormhole artifacts are notable in the b5000 neural DWI image (Fig2.b) due to the imperfect phase correction using filter-based phase correction. DLPC could retain high-frequency phase information and minimize wormhole artifacts while reducing noise floor (Fig2.c,d). Synthetic diffusion is frequently used, but its quality highly depends on ADC and acquired image quality. In non-endorectal prostate imaging at 1.5T, the rectified noise floor and wormhole artifacts are notable in the acquired b800 (Fig3.b) images, resulting in poor ADC and synthesized DWI images (Fig3.c,d). When the noise floor and wormhole artifacts were robustly removed using DLPC and deep learning reconstruction, both the acquired and synthesized DWI images showed improved diffusion contrast. DLPC and deep learning reconstruction also enable robust b2000 prostate diffusion imaging at 1.5T and 3T without endorectal coils. For healthy volunteers, the diffusion signals of the prostate quickly decay to the rectified noise floor at b2000 (Fig4.a,c). DLPC with deep learning reconstruction effectively reduced the noise floor and recovered the DWI signals, generating better contrast and conspicuous gland tissues at b2000 (Fig4.b,d). Parallel imaging could be used in reduced FOV DWI to further reduce the echo train length and geometric distortion, but it also further amplifies the noise in the center of prostate images (Fig5.a-c). With DLPC and deep learning reconstruction, the noise is well suppressed (Fig5.d-f), generating better contrast and conspicuity.

Conclusion

Compared to the filter-based phase correction, the deep-learning-based phase correction with deep-learning reconstruction could effectively reduce noise floor and wormhole artifacts, resulting in improved diffusion contrast. It can also improve the image quality of the synthesized DWI images due to a more accurate estimation of ADC.

Acknowledgements

No acknowledgement found.