Radiomics

Radiomics operates on a standardized pipeline: image acquisition, segmentation, feature extraction, and modeling. The process begins with high-quality, reproducible imaging, often using CT or MRI. Segmentation—delineating the region of interest (e.g., a tumor)—can be manual, semi-automated, or fully automated using deep learning. Feature extraction then computes hundreds of quantitative descriptors, including shape, intensity histograms, texture, and wavelet-based features.²

These features are subsequently reduced and selected to avoid overfitting, then fed into machine learning models to predict clinical outcomes such as survival, metastasis, or pathological subtype. The entire workflow requires careful standardization to ensure reproducibility across scanners and institutions, a challenge addressed by initiatives like the Image Biomarker Standardisation Initiative (IBSI).³

Radiomics has found its most extensive application in oncology, where it aids in tumor characterization, staging, and treatment planning. For example, radiomic signatures derived from CT images of non-small cell lung cancer can predict EGFR mutation status and overall survival, complementing genomic profiling.⁴ In breast cancer, MRI-based radiomics helps distinguish benign from malignant lesions and assess response to neoadjuvant chemotherapy.

Beyond oncology, radiomics is being explored in cardiology (e.g., characterizing myocardial fibrosis), neurology (e.g., identifying Alzheimer's disease patterns), and musculoskeletal imaging. The field also contributes to radiogenomics, linking imaging features to gene expression profiles, thereby offering a non-invasive window into tumor biology.⁵

Despite its promise, radiomics faces significant hurdles. Variability in image acquisition parameters, reconstruction algorithms, and segmentation methods can introduce non-biological noise, leading to poor reproducibility across centers. The IBSI has established guidelines for feature definition and reporting, but compliance remains inconsistent.³

Overfitting is another concern: with thousands of features and limited sample sizes, many published models may not generalize. The field is moving toward larger multi-center datasets, prospective validation, and transparent reporting (e.g., the Radiomics Quality Score). Additionally, the 'black-box' nature of deep learning models raises interpretability issues, prompting research into explainable AI for clinical trust.⁶

Radiomics extends beyond oncology into rare diseases and non-malignant conditions. For instance, it has been used to quantify liver fibrosis in chronic hepatitis, assess osteoporosis from routine CT scans, and even predict response to immunotherapy in melanoma—an area where conventional imaging often falls short.⁷

Historically, the concept of extracting quantitative imaging features predates the term 'radiomics' by decades, with early texture analysis in the 1970s. The term itself was coined in 2012 by Philippe Lambin and colleagues, who emphasized the need for 'high-throughput extraction of quantitative features.'¹ Another niche application is in veterinary medicine, where radiomics is being adapted for canine and feline tumors, potentially accelerating translation to human practice.

Texture analysis

A method of quantifying spatial variations in pixel intensity, used to capture heterogeneity within an image region.

Radiogenomics

The study linking imaging features to genomic data, aiming to infer genetic alterations from non-invasive images.

Image Biomarker Standardisation Initiative (IBSI)

An international collaboration that standardizes the definition and computation of radiomic features to improve reproducibility.