Fifty-eighth annual meeting of the american association of physicists in medicine
SU-F-R-57: Validation of Quantitative Radiomic Texture Features for Oncologic MRI: A Simulation Study
Radiomic texture features extracted from diverse MRI modalities have been investigated regarding their predictive and prognostic values in a variety of cancers. However, their validity has not yet been fully assessed. With the aid of a digital MRI phantom, the objective of this study was to examine the validity and reliability of MRI texture metrics.
MR signal of the employed digital phantom was simulated in a multiple coil setting with realistic acquisition noise. Three iterative algorithms based respectively on conjugate gradient (CG), total variation (TV), and wavelet regularization (WL) were used for image reconstruction. For each algorithm, 80 independent simulations were carried out with different levels of noise perturbation. 22 features related to grey-level co-occurrence matrices (GLCOM), zone size matrices (GLZSM), and neighborhood difference matrices (GLNDM) were evaluated for each resultant image within two ROIs featuring heterogeneous patterns of signal intensity on the ground truth image. Texture features extracted from these simulated images were compared to those from the ground truth image, and differences were identified.
In comparison to the ground truth data, texture features appearing on images reconstructed with CG, TV, and WL from signal with no noise perturbation varied with a range of 0.18–3.3×104%, 0.57–3.6×104%, and 0.84–3.5×104%, respectively while varying more significantly for noisy data with largest variation of 2.6×105% for CG, 2.2×105% for TV, and 2.8×105% for WL. Texture differences between ROIs also revealed considerable extent of variation from those on the ground truth image with a range of 1.91–4.9×104% for CG-based data, 27.50–3.0×104% for TV-based, and 11.99-3.2×104% for WL-based.
Variability of texture appearance on MRI with respect to the choice of reconstruction algorithm and noise level is significant and feature-dependent. Certain texture features may be preserved by MR imaging; however adequate precautions need to be taken on their validity and reliability.