https://atlas.rsna.org/schemas/2025-11/model.json

Reproducibility Analysis of Radiomic Features on T2-weighted MR Images in Neuroblastoma Tumors

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11294951/

1.0

European Union Horizon 2020 research and innovation programme, PRIMAGE project (SC1-DTH-07-2018), grant agreement no. 826494.

Retrospective, multicenter study with institutional ethics approvals at all sites; approved by the Ethics Committee for Investigation with Medicinal Products of the University and Polytechnic La Fe Hospital (ethics code 2018/0228); informed consent waived due to observational and retrospective design.

Radiomics feature extraction using PyRadiomics; automatic tumor segmentation using a deep learning nnU-Net convolutional neural network.

Automatic segmentation tool (nnU-Net for neuroblastoma MRI): https://github.com/lcerdaal/MRNeuroblastomaSegmentation/tree/main

Identifies stable and unstable radiomics features under common preprocessing and segmentation variations to inform robust biomarker development in pediatric neuroblastic tumors.

Research; technical validation of quantitative imaging biomarkers.

Excellent reproducibility defined as CCC ≥ 0.90; CoV categories: excellent ≤10%, good 11–20%, moderate 21–30%, poor >30%.

Automatic tumor detection/segmentation with nnU-Net, with radiologist visual validation and manual edits as needed; automated feature extraction via PyRadiomics.

Technical evaluation of radiomics feature reproducibility on T2/T2*-weighted MRI in pediatric patients with neuroblastic tumors at diagnosis and/or after initial chemotherapy, in a research environment.

T2/T2*-weighted MR images of primary neuroblastic tumors (abdominopelvic or cervicothoracic regions), with corresponding segmentation masks.

Reference pipeline: apply anisotropic diffusion denoising, N4 bias field correction, z-score normalization, and resampling (to 1×1×6 mm) prior to nnU-Net segmentation and PyRadiomics extraction of 107 features. Evaluate effects of alternative denoising, omission of inhomogeneity correction, omission of resampling or normalization, and mask erosion/dilation.

Heterogeneous multicenter dataset with varying acquisition parameters, time points (diagnosis vs post-chemotherapy), and tumor locations; only T2/T2*-weighted sequences analyzed; no scan–rescan repeatability available; analyses limited to neuroblastic tumors; potential unassessed confounders (e.g., vendor, sequence type) not stratified.

Report all preprocessing steps. For harmonized pipelines, the authors propose using anisotropic diffusion denoising, N4 bias field correction, z-score normalization, and resampling. Exercise caution with normalization due to its substantial impact on feature reproducibility.

Across perturbations, 57/93 (61%) features were highly reproducible for filtering changes; 41/93 (44%) for harmonization changes (resampling/normalization); 75/107 (70%) for mask modifications. All shape features remained stable under mask erosion/dilation.