Radiomics

Overview

Radiomics is the high-throughput extraction of a large number of quantitative imaging features from medical images (CT, MRI, PET) with the goal of capturing tumour phenotype non-invasively. Features include intensity statistics, shape descriptors, textural measures (GLCM, GLRLM, GLSZM, NGTDM), and wavelet-domain transforms. Extracted features are used in statistical or machine-learning models for prognosis, prediction, or diagnosis.

Cross-reference: ct-radiomics is the CT-specific instantiation; radiomic-signature-4-feature is the locked four-feature prognostic model from Aerts et al. (2014). For DICOM contour formats used as ROI inputs, see dicom-rt-struct.

Used by

• PMID:38362943 — RADCURE (3346-patient HNC RT planning CT dataset) is explicitly released to support radiomics and machine-learning research in head and neck radiation oncology; organ-at-risk and gross-tumor contours in DICOM RT-STRUCT format serve as the segmentation substrate for downstream radiomic feature extraction PMID:38362943.
• PMID:37397861 — RADCURE prognostic challenge benchmarked 12 ML models for 2-year OS prediction in 2,552 HNSC patients; both hand-engineered (PyRadiomics 1,316-feature) and deep-learning CT radiomics models were evaluated, with the finding that radiomic features added no prognostic value over EMR+tumor-volume; deep CNN approaches outperformed hand-crafted radiomics but neither beat simple clinical models PMID:37397861.

Notes

• Radiomics as a field was codified by Lambin et al. (2012) and Aerts et al. (2014 PMID:24892406); the latter demonstrated cross-disease prognostic transferability of a four-feature CT radiomic signature.
• Radiomic feature values are sensitive to image acquisition and reconstruction parameters; cross-institutional standardization remains an active challenge.

Sources

• PMID:38362943
• PMID:37397861

This page was processed by crosslinker on 2026-05-04.