IMG-71. Stability of radiomics in pediatric medulloblastoma: a multi-institutional preliminary analysis

Abstract BACKGROUND Acquiring MRI scans from multi-institutional clinical data presents challenges, particularly in ensuring stability of radiomic features across sites with different scanners and acquisitions. Feature stability is critical for subsequent analyses, including disease classification and development of biomarkers. This work evaluates the stability of radiomic features extracted from medulloblastoma (malignant pediatric brain tumor) across 3 sites with different tesla strengths. Our goal is to evaluate features’ stability prior to incorporating them into diagnostic/prognostic frameworks. METHODS T1-weighted MRI scans of 116 medulloblastoma patients were collected from Cincinnati Children’s Hospital Medical Center (Site1-n=40,3Tesla (T)), Children’s Hospital Los Angeles (Site2-n=44,1.5T), Children’s Hospital of Philadelphia (Site3-n=32,1.5T). Following preprocessing and segmenting the enhancing-lesion and edema sub-compartments, 43 textural features (Haralick, Laws, entropy) were extracted. The Coefficient of Variation (CV) was computed for each feature as follows: (EXP.1) Enhancing-lesion for Sites2,3 (1.5T); (EXP.2) Enhancing-lesion for Sites1,2,3 (1.5T,3T); (EXP.3) Edema for Sites2,3 (1.5T); (EXP.4) Edema for Sites1,2,3 (1.5T,3T). CV quantifies the relative variability of feature values, with a low value (<20%) indicating a stable feature and values <10% indicating high stability. RESULTS 20 features in EXP.1 exhibited CV values <10%, including, Haralick, correlation, entropy, sum-average, difference-variance, whereas 9 features exhibited values between 10%,20%. Similarly, 28 features in EXP.2 exhibited CV values <10%, including, Haralick, correlation, entropy, sum-average, whereas 16 features exhibited values between 10%,20%. For the edema region, 26 features in EXP.3 exhibited CV values <10%, including Haralick, correlation, entropy, whereas 14 features exhibited values between 10%,20%. When using all sites (EXP.4), 26 features exhibited CV values <10%, including, Haralick, contrast, sum-average, whereas 13 features exhibited values between 10%,20%. CONCLUSIONS Our preliminary, multi-institutional analysis for radiomics stability shows that our radiomic features exhibit limited variability across different sites and scanners, suggesting the resilience of features and the reliability of incorporating them in multi-site diagnostic/prognostic frameworks.