Standardizing Quality Assurance of Deep-Learning Diffusion-Weighted Imaging Using Qiba Profile and Phantom
Abstract
Purpose
Deep learning (DL) is increasingly used to reduce acquisition times, increase signal-to-noise ratio (SNR) and spatial resolution in magnetic resonance imaging (MRI). However, the impact on SNR and resolution is not measured in a quality assurance program. This study aims at evaluating the validity of DL reconstructed images in the context of diffusion-weighted imaging (DWI) with conformance to the QIBA DWI Profile metrics, for clinical and research package sequences in a quality assurance program.
Methods
A 1.5T system (MAGNETOM Sola, Siemens Healthineers AG, Forchheim, Germany) was used in this study with a clinical single-shot echo planar imaging (SSEPI) sequence and a research readout-segmented EPI (rsEPI). We evaluated the influence of DL on spatial resolution by computing the modulation transfer function (MTF) as well as QIBA DWI Profile metrics (SNR, ADC bias, b-value dependence, Random error, RC). The diffusion QIBA phantom was used. A student t-test was used.
Results
Our results show that the ADC bias percentage is increased by up to 4% and 7% for the central water vial when using DL with SSEPI and rsEPI sequence, respectively, compared to no DL. ADC bias can be higher for polyvinylpyrrolidone (PVP) concentration >40%. DL increases the maximum b-value dependence of ADC for all vials except one in both sequences. The SNR is increased, when using DL, by up to 25% and 100% for some PVP concentrations, and for all measured b-values in the case of SSEPI and rsEPI, respectively. For rsEPI, the SNR gain is statistically significant (p<0.05). For all b-values and both DWI sequences, the use of super-resolution significantly increases the spatial resolution as measured by the MTF.
Conclusion
DL increases the SNR and spatial resolution of DWI images at the expense of increased ADC bias and b-value dependence for both SSEPI and rsEPI DWI sequences.