Researchers at Beijing Tiantan Hospital employ a nested cross-validation radiomics pipeline with LASSO feature selection and TPOT-optimized random forest classifiers on contrast-enhanced T1-weighted MRI to noninvasively differentiate posterior pituitary tumors from pituitary neuroendocrine tumors and craniopharyngioma.
Key points
Extracted 1510 radiomic features from contrast-enhanced T1-weighted MRI, including shape, first-order, GLCM, GLRLM, GLSZM, and GLDM metrics.
Applied nested 10-fold cross-validation with LASSO-based dimensionality reduction and TPOT-optimized random forest classifiers to differentiate PPTs from NPPTs.
Achieved 0.786 accuracy, 0.818 AUC, 0.778 specificity, and 0.788 sensitivity in an independent validation cohort.
Why it matters:
Accurate noninvasive classification of pituitary tumors refines surgical planning, reduces intraoperative risks, and enhances patient outcomes.
Q&A
What is radiomics?
How does LASSO feature selection work?
Why use nested cross-validation and TPOT?
What clinical advantage does noninvasive tumor differentiation offer?
Read full article
Academy
Radiomics in Medical Imaging
Radiomics is a field that converts medical images into high-dimensional data through the extraction of quantitative features. It enables the identification of patterns and biomarkers within imaging data that may be imperceptible to human observers. By analyzing these features with machine learning algorithms, radiomics supports noninvasive diagnosis, prognostication, and treatment planning across various diseases.
Medical images such as CT, MRI, and PET scans contain rich information about tissue structure, texture, and intensity. Radiomic analysis involves the following steps:
- Image Acquisition and Standardization: High-quality, reproducible imaging protocols are essential. Variations in scanner settings and patient positioning are minimized to ensure consistent feature extraction.
- Segmentation: Regions of interest (ROIs), such as tumors or lesions, are delineated manually or automatically. Accurate segmentation is critical, as it defines the area from which features are computed.
- Feature Extraction: Radiomic features describe intensity distributions (first-order), spatial relationships (texture features such as GLCM, GLRLM, GLSZM, GLDM), shape descriptors, and transform-based features (wavelet, logarithm, exponential). This process can yield hundreds to thousands of quantitative descriptors.
- Feature Selection and Modeling: High-dimensional radiomic data are often pruned using algorithms like LASSO or tree-based methods to select the most informative features. Machine learning models—such as random forest, support vector machines, or neural networks—are then trained to perform classification, regression, or survival prediction tasks.
- Validation: Models are validated internally via cross-validation and externally on independent cohorts to assess robustness and generalizability. Nested cross-validation helps avoid overfitting by separating optimization from evaluation.
Radiomics has diverse applications in medical research and clinical practice. It has been used to:
- Differentiate tumor types and subtypes.
- Predict molecular and genetic profiles of tumors.
- Assess treatment response to chemotherapy and radiotherapy.
- Forecast patient prognosis and survival outcomes.
In aging and longevity research, radiomics can characterize tissue changes associated with age-related diseases, such as neurodegeneration and cardiovascular disorders. By linking imaging biomarkers to biological processes of aging, researchers can develop noninvasive tools for early detection, monitor disease progression, and personalize interventions that promote healthy lifespan.
Future Directions: Combining radiomics with genomics, proteomics, and clinical data within multi-omics frameworks promises more precise, personalized healthcare. Advances in deep learning-based feature extraction and federated learning may further enhance the reproducibility and scalability of radiomic analyses in multicenter studies.
