What Are Radiomics?

Radiomics is a field of medical research concerned with extracting large numbers of quantitative features from medical images and using statistical or machine learning methods to associate those features with clinical outcomes such as disease diagnosis, treatment response, and patient prognosis. The underlying premise is that the spatial distribution of pixel or voxel intensities in a CT, MRI, PET, or ultrasound image encodes biological information about tissue heterogeneity, morphology, and function that is not captured in the radiologist's qualitative report. By converting images into structured numerical data, radiomics attempts to make implicit imaging information explicit and reproducible.

The field emerged in the early 2010s from quantitative imaging research in oncology and draws on medical physics, medical image analysis, biostatistics, and machine learning. It is closely related to radiogenomics, which correlates radiomic features with genomic data, and to pathomics, which applies the same approach to histopathology slides. The term "radiomics" was formalized in a widely cited 2012 framework paper and has since expanded to encompass deep learning-derived features that complement or replace hand-crafted feature extraction.

Feature Extraction from Medical Images

The radiomic pipeline begins with image acquisition according to standardized protocols, followed by segmentation of the region of interest, typically a tumor, organ, or tissue lesion, and then extraction of features from the segmented volume. Feature categories include shape descriptors (volume, surface area, sphericity, compactness), first-order intensity statistics (mean, variance, skewness of the histogram), and texture features computed from matrices such as the gray-level co-occurrence matrix (GLCM), gray-level run length matrix (GLRLM), and gray-level size zone matrix (GLSZM). A single segmented tumor may yield hundreds to thousands of such features. As reviewed in PMC research on machine learning methods for radiomics, the high dimensionality of the feature space relative to the typical sample sizes in clinical studies creates a substantial risk of overfitting, which motivates feature selection and dimensionality reduction as critical intermediate steps.

Machine Learning and Predictive Modeling

After feature extraction and selection, machine learning models link radiomic features to clinical endpoints. Classical methods include logistic regression, support vector machines, and random forests, while deep learning convolutional neural networks learn features directly from image patches rather than from a pre-defined feature library. The PMC review on deep learning with radiomics describes how end-to-end deep learning architectures can improve predictive performance by capturing spatial context that hand-crafted features miss, at the cost of requiring larger training datasets and producing less interpretable models. Ensemble approaches that combine radiomics features with clinical variables, pathological grades, and genomic markers have shown better predictive accuracy than any single data type alone, particularly for endpoints such as overall survival and treatment response in lung, head and neck, and breast cancers.

Clinical Validation and Reproducibility

A persistent challenge in radiomics is the sensitivity of extracted features to acquisition parameters such as scanner manufacturer, reconstruction kernel, slice thickness, and contrast injection protocol, all of which can cause the same tissue to produce different feature values across sites. Multicenter validation studies and standardization initiatives, including the Image Biomarker Standardization Initiative (IBSI), have established reference definitions for feature calculation and digital phantom tests for software verification. The PMC overview of radiomics applications in oncology identifies reproducibility and prospective clinical validation as the primary barriers between research demonstrations and routine clinical implementation, noting that most published studies are retrospective and single-center.

Applications

Radiomics has applications across a range of clinical fields, including:

• Oncology, for non-invasive tumor subtype classification and prediction of treatment response to chemotherapy or radiotherapy
• Lung cancer screening, for distinguishing malignant nodules from benign findings in low-dose CT scans
• Precision medicine, for identifying imaging biomarkers that correlate with genomic mutations or immunotherapy outcomes
• Neurology, for characterizing brain tumor grade and predicting survival in glioma patients
• Cardiology, for quantifying myocardial texture and predicting adverse cardiac events from cardiac MRI