Background:
Ferumoxytol, an ultrasmall superparamagnetic iron oxide nanoparticle, is gaining clinical interest as an alternative to gadolinium-based contrast agent for MRI [1]. Due to uptake by macrophages, ferumoxytol is also being investigated for imaging of tumor-associated macrophages in adult and pediatric solid tumors [2]. Radiomic analysis of ferumoxytol-enhanced MR images (FE-MRI), termed FE-MRI nano-radiomics, may provide quantitative imaging signatures reflecting tumor immune biology such as myeloid content, based on intratumoral contrast patterns not captured by conventional imaging metrics [3]. However, before establishing such imaging biomarkers, it is essential to evaluate the reproducibility and robustness of nano-radiomic features (RFs).

Hypothesis/Objective:
We aimed to evaluate the impact of imaging parameters and test–retest variability on the stability of FE-MRI-derived RFs.

Methods:
In vivo studies were performed in a SV40 transgenic mouse model of neuroblastoma. Mice developed spontaneous bilateral tumors in the adrenals starting at 2.5 months of age. Mice (n=5) underwent pre-contrast T2w-MRI followed by post-contrast T2w-FE-MRI at 2 days after i.v. administration of ferumoxytol (20 mg Fe/kg). Tumors were segmented on MR images and RFs were calculated separately for each tumor (left and right) and subsequently averaged for downstream analyses. RFs were extracted with and without filtering, yielding 1,284 total features. Test–retest reproducibility was evaluated using three repeated acquisitions per mouse under identical imaging conditions (echo time [TE] = 60 ms, number of excitations [NEX] = 4), with animals removed and repositioned between scans to capture realistic variability. To assess sensitivity to acquisition parameters, additional experiments were performed by varying TE (60 vs. 107 ms; 5 mice × 2 TEs × 3 scans, [NEX] = 4) and NEX (2 vs. 4; 5 mice × 2 NEX × 3 scans, [TE] = 60 ms). For TE and NEX analyses, mice were scanned continuously without repositioning to isolate the effects of acquisition settings. Thus, shape-related RFs were excluded to minimize confounding from segmentation variability, resulting in 1,116 features analyzed. Reproducibility and effect sizes were quantified using intraclass correlation coefficient (ICC) and paired Cohen’s d, respectively.

Results:
In the test–retest analysis, radiomic analysis of post-contrast FE-MRI demonstrated substantially higher reproducibility of RFs, with 32.6% of features showing excellent agreement (ICC > 0.9) compared to 8.3% in pre-contrast scans. This improvement was more pronounced for features extracted without image filtering. Original (non-filtered) features were generally more stable, with only ~9% exhibiting a sign inversion in contrast effects across repeated measurements (Fig 1B). Most original features showed consistent positive contrast effects, indicating higher values in FE-MRI compared to pre-contrast imaging.

In the NEX sensitivity analysis, FE-MRI NEX = 4 scans yielded a higher proportion of highly reproducible features than NEX = 2 (27.1% vs. 21.7%), particularly for original and Laplacian-filtered features. NEX-dependent effect sizes were generally stable across contrast conditions. However, higher NEX preserved the direction but reduced the magnitude of contrast effects, suggesting attenuated sensitivity to contrast-induced changes with signal averaging (Table 1).

In the TE sensitivity analysis, FE-MRI at TE = 107 ms yielded the highest proportion of highly reproducible features (ICC > 0.9: 63%), compared to TE = 60 ms (27.1%), while pre-contrast proportions were similar between TE conditions (47% vs. 42.5%). Pre-contrast scans showed a balanced distribution of TE-dependent effect sizes, whereas FE-MRI shifted toward higher feature values at TE = 60 ms. Overall, FE-MRI demonstrated reduced TE-related variability, reflected by more features with smaller absolute Cohen’s d values, suggesting improved robustness of RFs to TE variation (Fig 2).

Conclusion:
Ferumoxytol-enhanced MRI consistently improved both reproducibility and robustness of nano-radiomic features compared to non-contrast scans. However, variability in reproducibility and sensitivity to acquisition parameters across imaging conditions highlights the need for standardized protocols in quantitative radiomic analysis.

Author
Esther Ngan
Baylor College of Medicine