Variable Resolution Sampling and Deep Learning Image Recovery for Accelerated Multi-Spectral MRI Near Metal Implants

TL;DR

This approach can potentially enhance MRI examinations near metal implants by reducing scan times or enabling higher resolution, and further prospective studies are needed to assess clinical value.

Abstract

Purpose: This study presents a variable resolution (VR) sampling and deep learning reconstruction approach for multi-spectral MRI near metal implants, aiming to reduce scan times while maintaining image quality. Background: The rising use of metal implants has increased MRI scans affected by metal artifacts. Multi-spectral imaging (MSI) reduces these artifacts but sacrifices acquisition efficiency. Methods: This retrospective study on 1.5T MSI knee and hip data from patients with metal hardware used a novel spectral undersampling scheme to improve acquisition efficiency by ~40%. U-Net-based deep learning models were trained for reconstruction. Image quality was evaluated using SSIM, PSNR, and RESI metrics. Results: Deep learning reconstructions of undersampled VR data (DL-VR) showed significantly higher SSIM and PSNR values (p<0.001) compared to conventional reconstruction (CR-VR), with improved edge sharpness. Edge sharpness in DL-reconstructed images matched fully sampled references (p=0.5). Conclusion: This approach can potentially enhance MRI examinations near metal implants by reducing scan times or enabling higher resolution. Further prospective studies are needed to assess clinical value.

Figures (5)

• Figure 1: A novel variable resolution (VR) sampling strategy to accelerate multi-spectral imaging in the presence of metal artifacts. The approach alternates between two acquisition methods: (1) odd-numbered frequency bins utilize conventional parallel imaging with partial Fourier sampling, while (2) even-numbered bins collect only auto-calibration signal (ACS) data in the $k_y-k_z$ plane to nearly halve the scan time. The resulting low-resolution images from even-numbered bins are subsequently reconstructed to full resolution using deep learning, enabling efficient multi-spectral acquisition without compromising image quality.
• Figure 2: In vivo results from a knee subject showcasing three consecutive spectral bin images with the most signal, along with the final combined RSOS image for (a) reference, (b) conventionally reconstructed zero-replaced (CR-ZReplace), (c) variable resolution sampling with zero-filled lines (CR-VR), (d) single-joint deep learning inferred zero-replaced images (SJ-DL ZReplace), (e) single-joint deep learning inferred VR images (SJ-DL VR), and (f) multi-joint deep learning inferred VR images (MJ-DL VR). A zoomed-in section of the entire field of view is displayed in the last column to highlight variations near the implant.
• Figure 3: In vivo results from a hip subject showcasing three consecutive spectral bin images with the most signal, along with the final combined RSOS image for (a) reference, (b) conventionally reconstructed zero-replaced (CR-ZReplace), (c) variable resolution sampling with zero-filled lines (CR-VR), (d) single-joint deep learning inferred zero-replaced images (SJ-DL ZReplace), (e) single-joint deep learning inferred VR images (SJ-DL VR), and (f) multi-joint deep learning inferred VR images (MJ-DL VR). A zoomed-in section of the entire field of view is displayed in the last column to highlight variations near the implant.
• Figure 4: Box and whisker diagrams showing the spread of structural similarity (SSIM) indices comparing the bin-combined reference versus CR-VR image and the reference versus those inferred through deep learning across different subjects and slices using (a) hip, (b) knee, and (c) combined multi-joint datasets. P-value results from Mann-Whitney U-Tests are indicated for each comparison. CR: Conventional Reconstruction, SJ: Single-Joint, MJ: Multi-Joint, DL: Deep Learning, VR: Variable Resolution.
• Figure 5: (a) Images and graphs illustrating the calculation of the RESI for a hip case. The red line indicates the region where the intensity profile was calculated for the reference, VR, and inferred SJ-DL VR and MJ-DL VR images. Dotted blank lines in the graphs mark areas where pixel information along the analyzed edges was gathered to estimate local finite difference slopes, contributing to subsequent RESI calculations. (b) Box and whisker plot of the normalized sharpness (RESI) for each reconstruction method. RESI: Relative Edge Sharpness Index, CR: Conventional Reconstruction, SJ: Single-Joint, MJ: Multi-Joint, DL: Deep Learning, VR: Variable Resolution.