Assessing the Reproducibility of MRI Radiomics for Predicting PIB PET Suvr
Abstract
Purpose
MRI-based radiomics can provide a PET-free way to estimate amyloid burden using MRI data alone. However, variability in radiomics feature stability and model performance poses technical challenges. This study examines the stability, reliability, and clinical utility of MRI-based radiomics for predicting global PiB PET SUVR expressed on the Centiloid scale.
Methods
MRI and paired PiB PET scans (GAAIN Centiloid 50–70 min) were preprocessed and segmented to extract cortical and cerebellar gray matter regions, followed by radiomics feature extraction. Feature stability was assessed using 300 bootstrap iterations with Random Forest importance ranking, retaining only consistently stable features. Model performance was evaluated using nested out-of-fold cross-validation and independent hold-out testing to ensure reproducibility and prevent data leakage, and quantified using R2 and MAE in SUVR units.
Results
Bootstrap analysis revealed instability in naive feature selection, with only a small subset of radiomic features consistently ranked as important. Stable features were primarily cortical texture and shape metrics (e.g., sphericity, gray-level non-uniformity, short-run high gray-level emphasis), whereas cerebellar features showed lower stability. After stability-based filtering, 17 robust MRI radiomics features were retained. Using these features, the Random Forest model achieved consistent and strong performance across validation strategies (R2 = 0.82, MAE = 0.10 SUVR). Error analysis indicated better performance at low SUVR levels and increased variability at higher amyloid burden.
Conclusion
Reproducible MRI-based prediction of PiB PET SUVR is achievable when radiomics feature stability is explicitly addressed. Stability-driven feature selection improves robustness by retaining consistently informative cortical radiomics features, enabling strong agreement with PiB PET SUVR using MRI data alone. Clinically, MRI-only SUVR estimation may be suited for early amyloid screening rather than precise quantification of advanced disease, underscoring the importance of reproducible radiomics pipelines, in line with prior radiomics literature highlighting the critical role of feature stability in quantitative imaging models.