ROI Dilation Disrupts Pediatric Brain Tumor ADC Radiomics

Key Takeaways

• In pediatric pre-surgical DWI for brain tumors, simulated ROI dilation was associated with more instability than erosion across first-order ADC radiomics, affecting 18 of 19 features after one iteration.
• Pilocytic astrocytomas showed the clearest susceptibility to boundary expansion, with more visibly altered average ADC intensity profiles than medulloblastomas or ependymomas.
• Matched training and evaluation with ground-truth contours outperformed matched perturbed contours, and a single dilation reduced classification accuracy more than a single erosion.
• Training across ROI variants and restricting analysis to stable features improved robustness to segmentation mismatch, although the mismatch effect was not eliminated.

Defining pediatric brain tumor margins on diffusion-weighted magnetic resonance imaging (DWI) can be difficult when low signal-to-noise, susceptibility artifacts, patient motion, and limited segmentation guidance make the border between solid tumor and surrounding tissue appear unstable from slice to slice. Because apparent diffusion coefficient (ADC) radiomics depends on which voxels are included before any model is built, small shifts in contour placement can change the measurement itself rather than only its interpretation. Investigators compared conservative and extensive boundary choices to see how those differences altered ADC features and downstream diagnostic classification.

In the Mulvany et al. PLOS One study of ROI segmentation and ADC radiomics in pediatric brain tumors, investigators used a retrospective single-center cohort from Birmingham Children’s Hospital’s Imaging of Tumours study, with pre-surgical DWI obtained before treatment. They identified 131 patients, successfully annotated 109, and included 106 in the full radiomics analysis. Diagnostic modeling was then restricted to a 69-patient subset with pilocytic astrocytomas, medulloblastomas, and ependymomas.

Ground-truth (GT) region of interest (ROI) contours were manually drawn on b0 images with b1000 as the main reference and T1- and T2-weighted magnetic resonance imaging as additional references for ambiguous regions, while limiting contours to solid tumor and excluding edema and cystic regions. Investigators then applied up to three 2D erosions or dilations with 3x3 all-ones kernels to simulate conservative and extensive boundary choices.

They extracted 19 first-order PyRadiomics features from ADC maps after isotropic 1 mm resampling, excluded textural features because of acquisition heterogeneity and large slice gaps, and used random forest (RF) classifiers with leave-one-out cross-validation (LOOCV) repeated 100 times.

Across successive perturbations, erosion and dilation both moved feature values away from GT, but dilation produced larger departures after one iteration across nearly every first-order feature. The effect was most evident in pilocytic astrocytomas, where average ADC intensity profiles changed more visibly with expanded boundaries than with conservative shrinkage and more than in medulloblastomas or ependymomas. Kurtosis was the lone paired comparison without a significant erosion-versus-dilation difference.

Matched training and evaluation with GT contours performed best, and within matched perturbed contours the larger accuracy penalty came from dilation rather than erosion. When models were trained on GT features and then evaluated on perturbed contours, accuracy fell by 3.8 ± 0.8% with E × 1 evaluation and 5.6 ± 0.9% with D × 1 evaluation.

Training across ROI variants and then restricting the model to median, mean, 10th percentile, root-mean-square, and 90th percentile features lowered those drops to 1.4 ± 0.7% and 2.9 ± 0.3%, although a residual mismatch effect remained.

This retrospective analysis came from a single UK pediatric center in Birmingham, so it should not be assumed to reflect U.S. or broader North American segmentation workflows. The erosion-and-dilation approach simulated annotation bias rather than directly measuring inter-reader variability, and dilated ROIs could extend into clearly erroneous cystic or cerebrospinal fluid regions. Diagnostic modeling covered only pilocytic astrocytomas, medulloblastomas, and ependymomas because other diagnoses were too sparse for stable training, and the findings were limited to first-order ADC radiomics for diagnostic classification rather than prognostic or treatment-planning tasks.

Overall, accurately drawn GT contours had the highest diagnostic utility in this study design, while extensive boundaries were more destabilizing than conservative ones for first-order ADC radiomics. Augmentation with stable-feature selection improved robustness but did not fully remove the effect of segmentation mismatch.

Clinician Questions

Which pediatric brain tumor types were included in the ADC radiomics classification models?

The diagnostic modeling results applied only to pilocytic astrocytomas, medulloblastomas, and ependymomas. The full cohort contained 25 diagnoses, but the other groups were too sparse for stable model training, which limits how far the accuracy findings extend across pediatric brain tumor subtypes.

Why did the analysis exclude textural ADC radiomic features in pediatric brain tumors?

The researchers limited the analysis to first-order ADC features because magnetic resonance imaging acquisition parameters were highly variable and the DWI and ADC images had large slice gaps. The paper also described first-order features as the conventional focus for diagnostic classification in this setting because they are more interpretable and available in standard clinical reporting tools.

How closely does simulated ROI erosion and dilation reflect real reader variability in pediatric brain tumor MRI segmentation?

The perturbations were designed to simulate conservative and extensive contouring bias, but they did not directly measure human inter-reader behavior. The altered contours followed GT geometry rather than image-driven human contouring, and dilated ROIs could include clearly erroneous cystic or cerebrospinal fluid regions, which is why the authors positioned this as a simulation rather than a direct reader-variability study.

Do these pediatric brain tumor ROI findings apply to prognosis or treatment-planning models?

The reported results were limited to diagnostic classification using first-order ADC radiomics. The authors did not test prognostic or treatment-planning tasks and noted that more extensive boundaries might still capture infiltration, edema, or cystic activity that could matter for other downstream applications.