RADIOMICS

RADIOMICS extracts and models quantitative radiomic features from fused FDG-PET (Fluorodeoxyglucose Positron Emission Tomography) and MRI scans to predict lung metastasis risk in soft-tissue sarcomas (STSs).

Key Features:

• Implementation: Implemented in MATLAB for radiological analysis, feature extraction, and modeling.
• Multimodal integration: Integrates FDG-PET and MRI, including T1-weighted and T2-weighted fat-suppression (T2FS) sequences, for combined analysis.
• Fusion method: Uses wavelet transform techniques to fuse FDG-PET with MRI and create composite textures.
• Feature extraction: Extracts nine non-texture features (including SUV metrics and shape descriptors) and forty-one texture features from separate and fused imaging modalities.
• Composite textures: Constructs composite texture representations from fused FDG-PET/MRI data to enhance texture-based characterization.
• Predictive modeling: Employs multivariable logistic regression for prediction of lung metastasis risk.
• Resampling strategy: Applies imbalance-adjusted bootstrap resampling across four steps: feature set reduction, feature selection, prediction performance estimation, and computation of model coefficients.
• Validation cohort: Evaluated on a cohort of 51 patients with histologically confirmed STSs using pre-treatment scans.
• Performance: The optimal four-texture-feature model achieved AUC 0.984 ± 0.002, sensitivity 0.955 ± 0.006, and specificity 0.926 ± 0.004 during bootstrap evaluation.

Scientific Applications:

• Metastasis risk prediction: Predicts lung metastasis risk in soft-tissue sarcomas from pre-treatment FDG-PET and MRI imaging.
• Early clinical evaluation: Enables early assessment of metastasis risk to inform clinical decision-making and treatment planning.
• Radiomic biomarker development: Supports development of multimodal radiomic biomarkers by combining FDG-PET and MRI texture features.

Methodology:

Composite textures are created by fusing FDG-PET and MRI (T1 and T2FS) using wavelet transform; nine non-texture (SUV metrics, shape descriptors) and forty-one texture features are extracted from separate and fused scans of 51 pre-treatment STS patients; multivariable logistic regression models are trained and evaluated using imbalance-adjusted bootstrap resampling with steps for feature set reduction, feature selection, prediction performance estimation, and computation of model coefficients.