Introduction to Radiomics

Abstract

Radiomics is a rapidly evolving field of research concerned with the extraction of quantitative metrics-the so-called radiomic features-within medical images. Radiomic features capture tissue and lesion characteristics such as heterogeneity and shape and may, alone or in combination with demographic, histologic, genomic, or proteomic data, be used for clinical problem solving. The goal of this continuing education article is to provide an introduction to the field, covering the basic radiomics workflow: feature calculation and selection, dimensionality reduction, and data processing. Potential clinical applications in nuclear medicine that include PET radiomics-based prediction of treatment response and survival will be discussed. Current limitations of radiomics, such as sensitivity to acquisition parameter variations, and common pitfalls will also be covered.

Keywords: PET; artificial intelligence; machine learning; radiomics; single-photon emission tomography.

FIGURE 1.

Visual representation of radiomic features: ¹⁸F-FDG PET and contrast-enhanced CT images of partly necrotic lung cancer of left lower lobe. Radiomic feature maps generated by moving small rectangular window over PET image, and calculating feature value for each position, reflect different aspects of glucose metabolism heterogeneity across tumor. Each feature map depicts a single radiomic feature, with high values corresponding to high signal intensities on gray-level feature map. Color-coded feature maps may be used for better visualization and as color overlay for CT. CE = contrast-enhanced; HH = high–high, or high-pass filtering in both directions.

FIGURE 2.

Calculation of radiomic texture features. Whereas GLCM relies on pixel pairs (here, interpixel distance = 0), GLRLM relies on runs, and GLSZM relies on areas of neighboring pixels with same gray-level.

FIGURE 3.

Radiomics workflow. First, ROI is defined or lesions are segmented. For ROIs or lesions, frequently a large number of feature candidates are extracted. Features that either represent variability in data most efficiently or serve a particular prediction model best are selected. Instead of selecting from predefined set of features, deep learning approaches link feature construction and modeling directly to further improve prediction accuracy and reliability.