Diffusion Weighted MRI Measures on a Commercially Available Low-Field (0.55T), Wide-Bore MRI Simulation System Using Deep Learning Reconstruction
Abstract
Purpose
Commercial availability of low-field (0.55 T), wide-bore (80 cm) MRI systems offer unique opportunities for treatment planning and response assessment in radiation oncology, especially when coupled with deep learning (DL) reconstruction algorithms. Understanding the bias and variance of quantitative diffusion measurements on these systems is important for implementing advanced planning protocols. In this study, diffusion weighted imaging (DWI) was performed with and without DL reconstructions to assess bias and variance of the apparent diffusion coefficient (ADC) on a low-field system at two time points.
Methods
ADC measurements were calculated from DW images of a NIST/RSNA diffusion phantom (Caliber MRI, Boulder, CO) consisting of 13 compartments with ADC values ranging from approx. 0.238 to 1.111x10-3 mm2/s at 0°C. A commercially available 0.55T (Siemens Free.Max, Germany) MRI system was used to acquire multiple b-value (N=4) DWI images at two time points (baseline and 2 weeks) with traditional and DL-based (DeepResolve; strength=8) reconstructions. Correlation and Bland-Altman repeatability (same reconstruction method) and agreement (different reconstruction method) measures were computed.
Results
Correlations of the traditional and DL-based ADC measures were significant for the two reconstruction algorithms (R2>0.9997; slopes ranging from 1.0073-1.0185). Mean ADC differences (agreement) and limits of agreement for the analysis of reconstruction method agreement ranged from -11.38 to -5.78x10-3 mm2/s and -27.84 to 7.53x10-3 mm2/s, respectively. BA analysis results for repeatability ranged from -38.66 to -29.16x10-3 mm2/s with limits of agreement ranging from -111.73-34.41 x10-3 mm2/s. No trends were observed in the BA analyses for either repeatability or reconstruction method agreement.
Conclusion
The clinical low-field, wide-bore MRI scanner evaluated in this study demonstrates minimal bias/variance of ADC measurements using traditional and DL-based reconstruction methods. Future work will focus on developing/optimizing quantitative DWI protocols on this system for in vivo assessment and prediction of treatment response.