1 Definition and scope

Radiomics is a quantitative approach to medical image analysis that converts routinely acquired scans into numerical data. By measuring patterns of intensity, shape, texture, and related properties, it seeks to uncover information that may not be apparent through visual inspection alone. The method is used to study disease biology, support diagnosis, estimate prognosis, and assess how lesions respond to therapy.

1.1 Etymology

The term radiomics combines “radiology” and the suffix “-omics,” which is used for fields that study large-scale datasets, such as genomics and proteomics. The name reflects the goal of treating images as a source of high-dimensional information rather than only as pictures for interpretation.

1.2 Core concept

The central idea of radiomics is that medical images contain measurable patterns linked to underlying tissue characteristics. These patterns can be extracted computationally from regions of interest and converted into features for analysis. In practice, the resulting data may help distinguish disease subtypes, quantify heterogeneity, or identify imaging signatures associated with outcomes.

1.3 Relationship to medical imaging

Radiomics depends on medical imaging because its input comes from standard clinical scans. Unlike conventional interpretation, which relies largely on visual assessment, radiomics adds a layer of numerical analysis. It does not replace radiology; instead, it extends imaging by enabling quantitative study of structures already visible on CT, MRI, PET, and related modalities.

1.4 Relationship to machine learning

Machine learning is often used to analyze radiomic features and build predictive models. These algorithms can identify complex relationships among variables that may be difficult to detect with simple rule-based methods. In many studies, radiomics serves as the feature-extraction stage, while machine learning provides classification, prediction, or clustering.

2 History and development

Radiomics developed from earlier efforts to quantify images in medicine. As computing power increased and digital imaging became routine, researchers gained the ability to extract and analyze many image-derived measurements at scale. The field has since expanded through the combined influence of statistics, informatics, and artificial intelligence.

2.1 Early quantitative imaging

Before the term radiomics became common, investigators were already measuring image-based variables such as lesion size, density, and signal characteristics. These early methods were often limited to a small set of manually defined descriptors. They laid the groundwork for more comprehensive feature extraction by showing that imaging could support numerical analysis.

2.2 Emergence of radiomics

Radiomics emerged as a distinct concept when researchers began describing imaging as a source of large, high-throughput data. This shift encouraged the extraction of hundreds or thousands of features from a single scan. The approach gained attention in oncology, where tumor heterogeneity and treatment response often vary in ways that images may capture.

2.3 Growth with artificial intelligence

The rise of machine learning and deep learning accelerated radiomics research. Improved algorithms made it easier to process large datasets, automate segmentation, and build predictive models. At the same time, broader access to digital archives and computational tools allowed studies to scale beyond small pilot projects.

Current work emphasizes reproducibility, multi-institution validation, and integration with clinical data. Researchers are also exploring hybrid systems that combine radiomic features with genomic, pathological, or laboratory information. Another active area is the use of automated pipelines that reduce manual intervention and improve consistency.

3 Imaging modalities used in radiomics

Radiomics can be applied to several imaging techniques, each contributing different kinds of information. The choice of modality depends on the clinical question, the tissue being examined, and the available acquisition protocol. In many studies, the same disease is evaluated across multiple modalities to capture complementary features.

3.1 Computed tomography

Computed tomography is widely used in radiomics because it produces standardized, high-resolution images with measurable attenuation values. CT is especially common in cancer research, where it can describe lesion size, density, margins, and internal heterogeneity. Its broad clinical use also makes it suitable for retrospective studies.

3.2 Magnetic resonance imaging

Magnetic resonance imaging offers strong soft-tissue contrast and multiple image contrasts, such as T1-weighted, T2-weighted, and diffusion-based sequences. Radiomic analysis of MRI can reveal structural and textural differences within organs and lesions. Because MRI protocols vary considerably, careful standardization is often necessary.

3.3 Positron emission tomography

Positron emission tomography provides metabolic and functional information rather than only anatomy. In radiomics, PET features may reflect tracer uptake patterns that are related to tumor activity or tissue physiology. PET is frequently paired with CT or MRI to combine metabolic and structural data.

3.4 Ultrasound

Ultrasound is attractive because it is widely available, inexpensive, and free of ionizing radiation. Radiomic studies using ultrasound examine echo patterns, boundaries, and internal texture. However, the method is sensitive to operator technique and acquisition settings, which can affect feature stability.

3.5 Multimodal imaging

Multimodal imaging combines information from two or more sources, such as PET-CT or MRI with pathology-derived measures. This approach can improve characterization by merging complementary feature sets. It also increases complexity, since images must be aligned, standardized, and analyzed in a coordinated manner.

4 Radiomics workflow

A radiomics study usually follows a structured pipeline from image acquisition to model evaluation. Each step can influence the final result, so methodological choices are important. Small variations in preprocessing or segmentation may alter extracted features and affect downstream predictions.

4.1 Image acquisition

The workflow begins with image acquisition under a defined protocol. Scanner type, reconstruction settings, slice thickness, pulse sequence, and timing can all influence the images. Consistent acquisition improves feature reliability and supports comparison across patients or sites.

4.2 Image preprocessing

Preprocessing prepares images for analysis by reducing technical variation. Common steps include resampling to a uniform voxel size, intensity normalization, noise reduction, and correction for artifacts. The goal is to make extracted features more comparable across scans.

4.3 Segmentation

Segmentation identifies the region of interest from which features will be measured. It may be performed manually, semi-automatically, or automatically. Because radiomic values depend on the selected boundary, segmentation quality is a major determinant of study accuracy.

4.4 Feature extraction

Feature extraction converts image data into numerical descriptors. Hundreds of variables may be computed from a single lesion, spanning basic intensity measures, geometric properties, and more complex texture patterns. These features form the input for statistical analysis or predictive modeling.

4.4.1 First-order features

First-order features describe the distribution of voxel intensities without considering spatial relationships. Examples include mean, median, variance, minimum, maximum, and skewness. They summarize how bright or dark a region appears overall.

4.4.2 Shape features

Shape features quantify the geometry of the segmented region. They may include volume, surface area, compactness, sphericity, and irregularity. Such measurements help characterize lesion form and extent.

4.4.3 Texture features

Texture features assess how intensities are arranged relative to one another. They can capture patterns such as homogeneity, contrast, granularity, and coarseness. These descriptors are often used to measure heterogeneity within tumors or organs.

4.4.4 Transform-based features

Transform-based features are derived after applying mathematical transforms such as wavelet, Laplacian, or filter operations. These methods reveal information at different spatial scales or frequency ranges. They are useful for identifying subtle structures that may not be obvious in the original image.

4.5 Feature selection

Feature selection reduces the number of variables used in a model. This step helps remove redundant or unstable features and limits noise. It also lowers the risk of overfitting when the dataset is small relative to the number of measurements.

4.6 Model building

Model building uses selected features to create a predictive or descriptive algorithm. Depending on the question, the model may classify disease, estimate survival, or predict therapy response. The modeling stage often combines imaging features with demographic, clinical, or laboratory data.

4.7 Validation and testing

Validation and testing determine whether a model performs well beyond the data used to train it. Internal validation may use cross-validation or bootstrapping, while external validation tests the model in an independent cohort. Strong validation is essential for judging real-world usefulness.

5 Types of radiomic features

Radiomic features are usually grouped by the type of information they capture. Some describe basic signal properties, while others summarize geometry or spatial relationships. More complex features may incorporate time, transforms, or interactions across scales.

5.1 Intensity-based features

Intensity-based features summarize the numerical values of pixels or voxels in the region of interest. They are useful for describing the overall signal level and spread. These features are often among the simplest and most reproducible.

5.2 Morphological features

Morphological features describe the outline and spatial form of an object. They are frequently used to assess whether a lesion has smooth, round, lobulated, or irregular boundaries. Shape descriptors can support differentiation between lesion types.

5.3 Textural features

Textural features quantify local or global patterns in image intensity. They are especially valuable for detecting heterogeneity, which may reflect mixed tissue composition, necrosis, fibrosis, or varying cellularity. Many radiomics models rely heavily on texture measures.

5.4 Temporal features

Temporal features are derived from changes over time, such as alterations across follow-up scans or dynamic imaging series. They can describe growth, shrinkage, or evolving signal patterns. These features are particularly relevant in treatment monitoring and longitudinal studies.

5.5 Higher-order features

Higher-order features are computed after applying filters, transforms, or more advanced mathematical operations. They may represent patterns at different resolutions or in transformed spaces. Such features can increase sensitivity to subtle image characteristics, though they may also be less stable.

6 Clinical applications

Radiomics has attracted attention because it may help extract clinically useful information from scans already obtained in routine care. Most applications have been studied in oncology, but the approach is expanding into other medical domains. Its practical value depends on the quality of the model and the relevance of the chosen features.

6.1 Cancer diagnosis

Radiomics can assist in identifying malignant lesions or distinguishing them from benign findings. In some studies, imaging features are used to support classification when conventional appearance is ambiguous. The method is not a standalone diagnostic test, but it can complement expert interpretation.

6.2 Tumor characterization

Radiomic analysis can describe properties such as aggressiveness, heterogeneity, and likely tissue composition. These measurements may help infer biological behavior from imaging alone. Tumor characterization is one of the most active areas of radiomics research.

6.3 Prognosis prediction

Some radiomic models aim to estimate outcomes such as survival, recurrence risk, or disease progression. By linking image patterns to clinical endpoints, researchers attempt to create noninvasive prognostic markers. These models are usually evaluated alongside established clinical predictors.

6.4 Treatment response assessment

Radiomics may help determine whether a tumor is responding to therapy before changes are obvious in size alone. Alterations in texture, density, or metabolic activity can sometimes precede anatomical shrinkage. This makes the approach useful in monitoring treatment effectiveness.

6.5 Disease recurrence prediction

After initial treatment, radiomics can be used to estimate the likelihood that disease will return. Features from baseline or early follow-up scans may contain signals related to residual disease or aggressive biology. Such predictions are being studied as a way to guide surveillance strategies.

6.6 Non-oncologic applications

Beyond cancer, radiomics is being explored in areas such as cardiovascular disease, lung disorders, neurologic conditions, and musculoskeletal imaging. In these settings, image-derived features may capture tissue remodeling, inflammation, or structural change. The field remains less mature outside oncology.

7 Data analysis methods

Radiomics relies on analytical methods that can handle many variables and often limited sample sizes. The selected method depends on the study goal, the type of outcome, and the amount of available data. Robust analysis is essential for generating models that are both accurate and interpretable.

7.1 Statistical modeling

Traditional statistical methods are used to assess associations between radiomic features and clinical outcomes. Regression models, survival analysis, and hypothesis testing remain important tools. They are often preferred when interpretability is a priority.

7.2 Machine learning approaches

Machine learning methods such as support vector machines, random forests, and ensemble models are widely used in radiomics. These algorithms can manage nonlinear relationships and interactions among features. Their performance, however, depends strongly on data quality and validation.

7.3 Deep learning integration

Deep learning can be integrated with radiomics in several ways. Neural networks may be used to automate segmentation, learn features directly from images, or combine extracted radiomic variables with raw imaging inputs. This integration can improve flexibility but may reduce transparency.

7.4 Feature reduction techniques

Feature reduction techniques simplify high-dimensional datasets before model training. Common methods include principal component analysis, correlation filtering, and regularization-based selection. These approaches help reduce redundancy and improve generalization.

8 Standardization and reproducibility

Standardization is a major concern in radiomics because image-derived features can change with technical conditions. Differences in scanners, protocols, or analysis software may lead to inconsistent results. Reproducibility is therefore a central issue in translating radiomics into clinical use.

8.1 Image protocol variation

Variation in acquisition protocols can alter feature values even when the underlying anatomy is similar. Differences in resolution, contrast timing, reconstruction algorithms, or scanner calibration may all matter. Harmonized protocols help reduce these effects.

8.2 Segmentation variability

Different observers or algorithms may draw slightly different boundaries around the same lesion. Such variability can change the resulting feature set. Studies often address this by using training protocols, consensus methods, or automated tools with quality control.

8.3 Feature harmonization

Feature harmonization aims to make measurements more comparable across sites and devices. Statistical techniques can adjust for batch effects or scanner-related differences. Harmonization is especially important in multicenter studies.

8.4 Reporting standards

Clear reporting standards improve transparency and allow other researchers to evaluate methods. Good reporting typically includes details on image acquisition, preprocessing, segmentation, feature definitions, model training, and validation. Incomplete descriptions make replication difficult.

8.5 Open-source tools and datasets

Open-source software and shared datasets support collaboration and reproducibility. They allow researchers to test methods under comparable conditions and to benchmark performance. Public resources also help train new investigators and accelerate method development.

9 Advantages and limitations

Radiomics offers a way to mine clinically acquired images for additional information without requiring new invasive procedures. At the same time, it faces methodological and practical challenges. Its value depends on careful study design, validation, and integration with existing clinical knowledge.

9.1 Benefits in clinical research

Radiomics can generate noninvasive biomarkers from data already present in routine care. It may detect subtle patterns that are difficult to judge visually and can support large retrospective studies. The approach is also compatible with repeated imaging, making longitudinal assessment possible.

9.2 Sources of bias

Bias may arise from limited sample size, nonrepresentative cohorts, inconsistent acquisition, or selective outcome reporting. If these issues are not controlled, the resulting model can appear stronger than it truly is. Balanced study design is therefore essential.

9.3 Overfitting and generalizability

Because radiomics datasets often contain many features relative to the number of patients, overfitting is a frequent risk. A model may perform well on training data but fail on new cases. External validation is the main safeguard against this problem.

9.4 Interpretability challenges

Some predictive models are difficult to interpret clinically, especially when they use many correlated features or complex algorithms. This can limit trust and complicate adoption. Researchers often try to balance performance with transparency.

9.5 Regulatory and implementation barriers

Before radiomics can be widely adopted, it must fit into clinical workflows and meet quality requirements. Regulatory review, software integration, and prospective evaluation may all be needed. Practical barriers also include data storage, interoperability, and training.

10 Future directions

The future of radiomics is likely to involve deeper integration with other biomedical data and more automated analysis pipelines. Progress will depend on standardization, validation, and clinically meaningful endpoints. As methods mature, the field may play a larger role in precision medicine.

10.1 Radiogenomics

Radiogenomics links imaging features with genomic or molecular profiles. The goal is to infer biological information from scans that corresponds to gene expression, mutations, or pathway activity. This area may help connect imaging phenotypes with tumor biology.

10.2 Multimodal biomarker development

Future biomarkers may combine radiomic features with laboratory, clinical, and molecular data. Multimodal models can capture complementary dimensions of disease. Such integration may improve accuracy compared with imaging alone.

10.3 Personalized medicine

Radiomics is often discussed as a tool for tailoring care to individual patients. By identifying imaging signatures associated with risk or treatment sensitivity, it may support more precise decisions. Realizing this promise will require robust clinical evidence.

10.4 Automated clinical decision support

Another likely direction is the incorporation of radiomic models into decision-support systems. These tools could assist clinicians by flagging risk patterns, suggesting follow-up intensity, or summarizing quantitative image findings. Successful use will depend on reliability, usability, and clear demonstration of benefit.