This study evaluates the stability of texture features in MRI reconstructed with and without a vendor-provided deep learning method across different acceleration techniques, including GRAPPA and SMS. The aim is to assess how such reconstructions influence the preservation of diagnostically relevant image textures in phantom and brain scans. MRI scans were performed on a 1.5T system using a fluid-filled phantom and healthy volunteers. T2-weighted and FLAIR sequences were acquired with varying GRAPPA, SMS, and Deep Resolve settings. A total of 570 texture features were extracted from regions of interest using MaZda software. Stability was evaluated using Lin’s concordance correlation coefficient and the Wilcoxon rank-sum test. In phantom data, high concordance between Deep Resolve and standard reconstruction was rare (0–2 features with Lin > 0.9), and the number of features with Lin > 0.8 decreased with higher acceleration. SMS2 and GRAPPA R = 2 preserved the largest number of features (up to 27–151 with Lin > 0.8). In brain images, 36–414 features achieved Lin > 0.8 depending on tissue and acceleration, with cerebrospinal fluid and white matter showing greater concordance than cortex. Deep Resolve also produced a higher number of features meeting the Wilcoxon criterion for equal medians (p ≥ 0.05), particularly under moderate acceleration (e.g., up to 486 features in cortex for GRAPPA R = 2). Deep Resolve modifies MRI texture features in a manner that depends on tissue type and acceleration factor. The highest texture stability was observed under moderate acceleration, while higher acceleration consistently reduced concordance. These findings highlight the need to account for reconstruction effects when using texture-based or radiomic analysis.
Obuchowicz et al. (Mon,) studied this question.