Radiomics Feature Extraction: A Practical Workflow for Quantitative Imaging Biomarkers

Radiomics feature extraction is the process of converting medical images into quantitative, mineable data by computing mathematical descriptors of intensity, shape, and texture within defined regions of interest. For researchers, students, analysts, and life-science professionals, the practical challenge is also running software but building a reproducible pipeline that produces features that can be compared across patients, scanners, and institutions. This article provides a step-by-step workflow covering image acquisition, segmentation, preprocessing, feature computation, and quality control, with emphasis on reproducibility and the reporting standards that determine whether a radiomics study can be externally validated.

What Radiomics Measures and Why It Matters

Radiomics is a rapidly evolving field concerned with the extraction of quantitative metrics, called radiomic features, from medical images. These features capture tissue and lesion characteristics such as heterogeneity and shape and may be used alone or in combination with demographic, histologic, genomic, or proteomic data for clinical problem solving. The fundamental principle is that biomedical images contain predictive information not discernible to the human eye that can be mined through quantitative image analysis.

The clinical rationale is straightforward. Routine imaging generates large volumes of anatomical and functional data, and radiomics aims to enhance the data available to clinicians through advanced mathematical analysis. By extracting the spatial distribution of signal intensities and pixel interrelationships, radiomics quantifies textural information using analysis methods from artificial intelligence. Applications span diagnosis, prognosis, treatment response prediction, and disease monitoring across multiple cancer types and beyond.

The field has matured to the point where dedicated software platforms exist. PyRadiomics, a flexible open-source platform implemented in Python, extracts a large panel of engineered features from medical images and can be used standalone or within 3D Slicer. The platform was developed to establish a reference standard for radiomic analyses and to address the lack of standardized algorithm definitions and image processing that severely hampers reproducibility and comparability of results.

The Radiomics Pipeline at a Glance

The complete radiomics workflow follows a defined sequence of stages, each with specific decisions that affect downstream feature values. The table below summarizes the core stages, the key decisions at each stage, and the primary reproducibility risks.

Image Acquisition and Protocol Standardization

Image acquisition is the first and most consequential stage of the radiomics pipeline. The images themselves are the raw material, and any variation introduced at acquisition propagates through every subsequent stage. Radiomic features are sensitive to acquisition parameter variations, and this sensitivity is one of the most commonly cited limitations of the field.

Modality-Specific Considerations

Computed tomography, magnetic resonance imaging, and positron emission tomography each present distinct challenges. For CT, the reconstruction kernel, tube current, slice thickness, and contrast administration timing all influence feature values. For MRI, the sequence type, echo time, repetition time, field strength, and coil selection affect intensity distributions. For PET, the reconstruction algorithm, matrix size, and smoothing filter impact feature computation.

The practical implication is that radiomics studies should use images acquired under consistent protocols whenever possible. When pooling data from multiple institutions, researchers must document acquisition parameters and consider harmonization techniques. The Image Biomarker Standardisation Initiative was founded to harmonize feature definitions and preprocessing, and it specifies the mathematics of radiomic features and upstream steps including interpolation, resegmentation, intensity discretization, convolutional filtering, and feature aggregation.

Phantom Validation

Phantom studies provide a controlled method for assessing how acquisition and preprocessing affect feature stability. In one CT phantom study using the Credence Cartridge Radiomics phantom, researchers extracted 51 radiomic features from 11 texture regions of interest and processed them with 19 preprocessing algorithms. The results showed that the preprocessing filter, the texture cartridge, and the feature category all affected the radiomic properties of the phantom images. Homogeneous intensity phantom inserts were less sensitive to preprocessing changes than regular directional honeycomb textures.

For MRI, phantom studies on 1.5T magnetic resonance-guided linear accelerator systems assessed test-retest repeatability, longitudinal repeatability, and inter-platform reproducibility. Optimal preprocessing strategies varied across imaging sequences and evaluation tasks, with robust features defined as those with intraclass correlation coefficient values above 0.9 and coefficient of variation values below 10 percent.

Segmentation: Defining the Region of Interest

Segmentation delineates the region of interest from which features will be extracted. This stage is a major source of variability because different observers and different algorithms produce different boundaries, and those differences propagate into feature values.

Manual Segmentation

Manual delineation by a radiologist or trained observer remains common, particularly in research settings. The advantages are clinical expertise and the ability to handle complex anatomy. The disadvantages are time consumption and inter-observer variability. Studies should report the number of observers, their experience level, and the method used to resolve disagreements, such as consensus review or averaging.

Semi-Automated and Automated Segmentation

Semi-automated methods combine human input with algorithmic assistance, such as region growing or thresholding. Fully automated methods use deep learning models to segment structures without human interaction. In a study of hypopharyngeal cancer on MRI, convolutional neural network architectures including U-Net and DeepLab V3+ were trained on 178 patient images and tested on 44 patient images. The DeepLab V3+ model achieved a dice similarity coefficient of 0.77 compared to 0.75 for U-Net, with the difference most pronounced for small tumor volumes below 10 cubic centimeters. For first-order feature extraction, both models exhibited high agreement with manual delineation, with intraclass correlation coefficients ranging from 0.71 to 0.91.

The practical lesson is that automated segmentation can produce reasonable results, but the choice of architecture matters, and validation against manual delineation is essential before relying on automated features.

Segmentation Quality Control

Regardless of the segmentation method, quality control is mandatory. Visual inspection of every segmentation against the source image is the minimum standard. Quantitative metrics such as the dice similarity coefficient, Jaccard index, and average surface distance provide objective measures of agreement between automated and manual segmentations. When segmentation quality is poor, features extracted from those regions will be unreliable, and downstream models will suffer.

Image Preprocessing and Its Impact on Feature Stability

Preprocessing transforms the raw image before feature computation. Common steps include voxel resampling, intensity normalization, intensity discretization, and filtering. The choices made at this stage have a substantial impact on feature reproducibility and on the performance of downstream classification models.

Resampling and Interpolation

Medical images have anisotropic voxel dimensions in many cases, meaning the slice thickness differs from the in-plane resolution. Resampling to isotropic voxels is often performed to make features rotationally invariant and comparable across datasets. The choice of interpolation method, such as linear or B-spline, affects feature values, particularly for texture features.

Intensity Normalization

Intensity normalization addresses differences in signal intensity ranges across images. Z-score normalization is a common approach that centers the intensity distribution at zero and scales it to unit variance. In a study of Parkinson's disease motor subtype analysis using 210 T1-weighted MRI scans, five preprocessing pipelines were applied and 22,560 radiomic features were extracted from 16 brain regions. The combination of bias field correction and Z-score normalization showed the most significant improvement in classification performance after excluding non-reproducible features, with the area under the receiver operating characteristic curve increasing from 0.49 to 0.64.

Intensity Discretization

Radiomic texture features require intensity values to be discretized into a finite number of bins. The two main approaches are fixed bin width and fixed bin number. The choice of bin width or bin number substantially affects feature values. In the MR-Linac phantom study, intensity discretization using bin width or bin number was evaluated, and the optimal strategy varied across imaging sequences and evaluation tasks.

Filtering

Filters enhance or suppress specific image characteristics before feature extraction. Common filters include Laplacian of Gaussian, wavelet transforms, and smoothing filters. In the CT phantom study, the Laplacian filter, log filter, resample, and bit depth rescale range preprocessing algorithms hugely influenced histogram and gray level co-occurrence matrix features.

Preprocessing and Feature Reproducibility

The evidence consistently shows that preprocessing choices determine which features are reproducible. In the Parkinson's disease study, wavelet-based features showed the highest reproducibility, with 37 percent demonstrating excellent intraclass correlation coefficient values of 0.90 or above. Texture-based features, particularly from the gray level co-occurrence matrix and gray level size zone matrix, were among the most reproducible across preprocessing methods. Excluding non-reproducible features generally improved classification performance.

In a study of non-small cell lung cancer using CT images of 107 tumors with volumes ranging from 5 to 567 cubic centimeters, features were calculated with no preprocessing, 8-bit depth resampling, Butterworth smoothing, or both resampling and smoothing. The preprocessing technique affected the volume dependence and prognostic potential of features in univariate analyses.

Feature Computation and Standardized Definitions

Feature computation converts the preprocessed image and segmentation into numerical values. The features themselves fall into several families, each capturing different aspects of the image.

Feature Families

First-order features describe the distribution of voxel intensities within the region of interest without considering spatial relationships. These include mean, median, standard deviation, skewness, kurtosis, and entropy. Shape features describe the geometric properties of the region, including volume, surface area, sphericity, and compactness. Texture features capture spatial arrangements of intensities and include the gray level co-occurrence matrix, gray level run length matrix, gray level size zone matrix, neighborhood gray tone difference matrix, and gray level dependence matrix.

The Image Biomarker Standardisation Initiative

The lack of standardized algorithm definitions and image processing severely hampers reproducibility and comparability of results. The Image Biomarker Standardisation Initiative addresses this by specifying the mathematics of radiomic features and upstream steps. It provides compliance resources including digital phantoms, benchmark values, and validation checklists. These measures improve cross-software agreement, especially for first-order and shape features, while clarifying residual variability in higher-order textures.

IBSI is embedded in widely used platforms including PyRadiomics and LIFEx, enabling more stable multicenter workflows and integration with machine learning models. For researchers, using IBSI-compliant software and reporting the IBSI version used is a concrete step toward reproducibility.

Software Platforms

PyRadiomics is a flexible open-source platform capable of extracting a large panel of engineered features from medical images. It is implemented in Python and can be used standalone or within 3D Slicer. Source code, documentation, and examples are publicly available. Other platforms include LIFEx and Cancer Imaging Phenomics Toolkit, known as CaPTk.

Conventional tools such as PyRadiomics and CaPTk rely on extensive handcrafted feature sets, which often result in redundancy and necessitate further optimization steps. Newer approaches aim to address this limitation. Spectral Entropic Radiomics Feature Extraction, known as SERFE, integrates spectral frequency decomposition, entropy-driven feature selection, and graph-based spatial encoding. In an evaluation using the public TCIA glioblastoma dataset, SERFE generated a refined feature set of 350 radiomic features from an initial pool of 2,260, achieving a 92 percent stability score and 91.7 percent classification accuracy.

Quality Control and Feature Stability Assessment

Quality control is not an optional step at the end of the pipeline. It is an integral part of every stage, and its absence is a primary reason why radiomics models fail to translate to clinical practice.

Test-Retest and Longitudinal Repeatability

Test-retest analysis involves imaging the same subject or phantom twice under identical conditions and comparing the extracted features. Features that vary substantially between repeated acquisitions are not reliable biomarkers. Longitudinal repeatability extends this concept to imaging performed over time, which is essential for treatment response assessment.

In the MR-Linac phantom study, the proportion of robust features in test-retest analysis reached 85.87 percent for T1-weighted images, 89.13 percent for T2-weighted images, and 90 percent for FLAIR images, though the full results varied by sequence and preprocessing strategy.

Intraclass Correlation Coefficient and Coefficient of Variation

The intraclass correlation coefficient and coefficient of variation are the standard metrics for feature stability. Features with intraclass correlation coefficient values above 0.9 and coefficient of variation values below 10 percent are generally considered robust. These thresholds should be reported explicitly, and the choice of threshold should be justified.

Phantom-Based Quality Control

Phantoms provide a standardized substrate for quality control. The Credence Cartridge Radiomics phantom contains multiple texture inserts that can be imaged repeatedly to assess feature stability across acquisitions and preprocessing pipelines. The American College of Radiology phantom serves a similar purpose for MRI systems.

Feature Selection and Dimensionality Reduction

Radiomics pipelines typically extract hundreds or thousands of features from each image. This high dimensionality creates problems for model building, including overfitting and the risk of finding spurious associations. Feature selection and dimensionality reduction are therefore essential stages.

The Curse of Dimensionality

With limited sample sizes, which are common in medical imaging studies, the number of features can exceed the number of patients by a wide margin. Models built on such data are prone to overfitting, meaning they perform well on the training data but fail on new data. The radiomics quality score literature consistently identifies limited sample sizes and lack of external validation as major impediments to clinical translation.

Selection Strategies

Feature selection methods fall into several categories. Filter methods rank features by statistical properties such as correlation with the outcome or reproducibility across test-retest data. Wrapper methods evaluate feature subsets by model performance. Embedded methods perform selection during model training, such as regularization techniques that shrink unimportant feature coefficients to zero.

The SERFE approach demonstrates an alternative strategy. By integrating spectral frequency decomposition, entropy-driven feature selection, and graph-based spatial encoding, it generated a refined feature set without requiring post-extraction feature reduction. This approach maintains computational efficiency while capturing tumor heterogeneity.

Volume Dependence

Some radiomic features are strongly correlated with lesion volume, which can confound analyses. In the non-small cell lung cancer study, features with very high volume correlations above 0.95 were identified, and normalization for volume was applied. Researchers should assess the volume dependence of their features and consider whether volume itself should be included as a covariate in models.

Model Building and Validation

The ultimate goal of radiomics feature extraction is to build predictive models that support clinical decision making. The modeling stage has its own set of methodological requirements.

Data Splitting and Cross-Validation

Models should be developed on a training dataset and evaluated on a separate validation dataset. Internal validation methods such as cross-validation provide an estimate of model performance, but they do not substitute for external validation on data from a different institution or acquisition protocol. The radiomics quality score explicitly rewards studies that perform external validation.

Performance Metrics

The choice of performance metrics depends on the prediction task. For binary classification, the area under the receiver operating characteristic curve is commonly reported. For survival analysis, the concordance index is used. Calibration, which assesses whether predicted probabilities match observed frequencies, is also important and frequently underreported.

Clinical Utility Assessment

Beyond statistical performance, models should be assessed for clinical utility. This includes considering the clinical workflow into which the model would be integrated, the actionable decisions it would inform, and the potential harms of incorrect predictions. A model with high area under the curve but no clear clinical application has limited value.

The Radiomics Quality Score and Reporting Standards

The radiomics quality score is a tool developed to assess the rigor of radiomics studies. It has been widely adopted by researchers and provides a structured framework for evaluating study quality across multiple dimensions.

Radiomics Quality Score Version 1.0

The original radiomics quality score established a benchmark for evaluating radiomics studies. It includes criteria covering image protocol quality, segmentation, preprocessing, feature selection, validation, and reporting. The maximum score is 36 points, and higher scores indicate higher methodological quality.

Evidence from Meta-Analyses

A systematic review and meta-analysis of radiomics quality score application included 130 systematic reviews and extracted 3,258 quality scores. The overall mean radiomics quality score was 9.4 with a standard deviation of 6.4, representing 26.1 percent of the maximum possible score. Quality scores were positively correlated with publication year, and scores were significantly higher after publication of the radiomics quality score itself. Only 7.2 percent of scores reached 50 percent or more of the maximum. Criteria scoring errors were discovered in 39.8 percent of articles that reported criteria scores.

A separate meta-analysis found that after more than a decade of investigation, issues with workflow standardization, model reproducibility, validation, and data accessibility persist and impede the clinical translation of radiomics-based models.

Radiomics Quality Score 2.0

An updated framework, radiomics quality score 2.0, was introduced to account for evolving knowledge and to ensure optimal evaluation of study quality through the inclusion of fairness, explainability, rigorous quality control, and harmonization. The updated score maintains the scientific rigor of its predecessor and introduces radiomics readiness levels, inspired by the technology readiness level framework. These readiness levels reflect nine distinct levels of incremental improvement in radiomics research with the ultimate aim of clinical implementation.

Practical Reporting Recommendations

For researchers building radiomics pipelines, the following reporting elements are essential. Report acquisition parameters for every imaging protocol. Describe the segmentation method and its validation. Specify all preprocessing steps including resampling, normalization, and discretization parameters. Name the feature extraction software and version. Report feature stability metrics. Describe feature selection methods. Provide the full model specification including all coefficients. Report internal and external validation results. State the radiomics quality score for the study.

Common Failure Patterns in Radiomics Studies

Understanding how radiomics studies fail is as important as knowing how to build them correctly. The meta-analytic evidence reveals recurring patterns.

Failure Pattern 1: Inconsistent Feature Definitions

Different software implementations and versions compute features differently. A feature named entropy in one platform may not match entropy in another. The Image Biomarker Standardisation Initiative was created specifically to address this problem, and studies that do not use standardized definitions cannot be compared with each other.

Failure Pattern 2: Inadequate Segmentation Validation

Automated segmentation models are often trained and tested on data from a single institution. When applied to data from other institutions with different scanners and protocols, performance degrades. Studies that do not validate segmentation quality on their own data risk extracting features from incorrect regions.

Failure Pattern 3: Ignoring Preprocessing Effects

The evidence is clear that preprocessing choices affect feature reproducibility and model performance. Studies that do not report preprocessing parameters or that use arbitrary preprocessing choices without justification produce results that cannot be replicated.

Failure Pattern 4: Overfitting with High-Dimensional Features

Extracting thousands of features from dozens of patients and building models without proper validation produces models that appear accurate in the training data but fail completely on new data. The radiomics quality score literature consistently identifies this as a major problem.

Failure Pattern 5: Incomplete Reporting

Studies that omit acquisition parameters, segmentation details, preprocessing steps, or validation results cannot be evaluated or replicated. The radiomics quality score provides a structured checklist, and its widespread adoption has improved reporting over time, though the overall quality remains low.

Failure Pattern 6: Lack of External Validation

Models validated only on the training data or on internal test sets may not generalize. External validation on data from different institutions, scanners, and patient populations is the gold standard, yet it remains rare in the literature.

Data Sharing and Reproducibility Infrastructure

Reproducibility extends beyond methodological reporting to data and code sharing. The broader biomedical research community has established frameworks that radiomics researchers should adopt.

FAIR Guiding Principles

The FAIR Guiding Principles describe the characteristics that data resources should possess to be findable, accessible, interoperable, and reusable. These principles apply directly to radiomics datasets, which should be deposited in repositories that support persistent identifiers, rich metadata, and clear usage licenses.

Genomic Data Sharing Policy

For radiogenomics studies that combine imaging features with genomic data, the NIH Genomic Data Sharing Policy provides requirements for data sharing, including informed consent, data anonymization, and deposition in approved repositories. Researchers combining radiomics with genomic data should review these requirements early in their study design.

Data Repositories

The National Center for Biotechnology Information provides access to a range of biomedical data resources, including genomic, transcriptomic, and imaging data. The Cancer Imaging Archive hosts many public radiomics datasets, including the glioblastoma dataset used in the SERFE evaluation. The European Bioinformatics Institute offers training resources for bioinformatics and data management.

Practical Implementation Steps

The following steps provide a concrete workflow for implementing a reproducible radiomics feature extraction pipeline.

Step 1: Define the Clinical Question and Study Design

Specify the clinical problem, the imaging modality, the patient population, and the outcome of interest. Determine the sample size needed for the planned analyses. Register the study protocol if appropriate.

Step 2: Standardize Image Acquisition

Document all acquisition parameters. If using retrospective data, assess the variability of acquisition parameters across the dataset. Consider whether harmonization techniques are needed.

Step 3: Develop and Validate Segmentation

Choose a segmentation approach appropriate for the anatomy and modality. If using automated segmentation, validate against manual delineation on a representative sample. Report dice similarity coefficient and other agreement metrics.

Step 4: Define the Preprocessing Pipeline

Select resampling, normalization, and discretization parameters based on the modality and the feature families of interest. Document all parameters. Consider testing multiple preprocessing strategies and assessing feature stability across them.

Step 5: Extract Features with Standardized Software

Use IBSI-compliant software such as PyRadiomics. Record the software version and all parameters. Export features in a structured format with feature names that match the IBSI definitions.

Step 6: Assess Feature Stability

Perform test-retest analysis if repeated acquisitions are available. Calculate intraclass correlation coefficient and coefficient of variation for each feature. Exclude features that do not meet stability thresholds.

Step 7: Perform Feature Selection and Modeling

Apply feature selection methods appropriate for the sample size. Build models with proper internal validation. If possible, perform external validation on an independent dataset.

Step 8: Report According to Radiomics Quality Score Standards

Complete the radiomics quality score checklist. Report all acquisition, segmentation, preprocessing, feature extraction, and modeling details. Deposit data and code in appropriate repositories.

Records and Measurements for Quality Assurance

Maintaining detailed records is essential for reproducibility and for diagnosing problems when they arise.

Acquisition Log

Record the scanner model, software version, acquisition protocol, reconstruction parameters, contrast agent and timing, and any protocol deviations for every image in the study.

Segmentation Log

Record the segmentation method, the observer or algorithm, the date, and the results of any quality checks. For manual segmentation, record the observer identity and experience. For automated segmentation, record the model version and the validation metrics.

Preprocessing and Feature Extraction Log

Record every preprocessing parameter and every feature extraction setting. Record the software version and any configuration files. This log should be sufficient for another researcher to reproduce the exact feature values from the raw images.

Model Development Log

Record the data splitting method, the feature selection approach, the model type and hyperparameters, and the validation results. Record the final model coefficients and the performance metrics on all validation sets.

Professional Escalation Criteria

Radiomics is a research tool, and its outputs should not be used for clinical decision making without appropriate oversight. The following criteria indicate when to escalate concerns to a supervisor, collaborator, or institutional review body.

Segmentation Quality Concerns

If automated segmentation produces dice similarity coefficients below acceptable thresholds, or if visual inspection reveals systematic errors, stop the pipeline and consult with an imaging expert before proceeding.

Feature Instability

If test-retest or longitudinal repeatability analysis shows that a large proportion of features are unstable, do not proceed to model building with those features. Consult with a medical physicist or imaging scientist to identify the source of variability.

Model Performance Concerns

If model performance differs substantially between training and validation sets, or if external validation fails, do not report the model as clinically useful. Consult with a biostatistician to assess whether the modeling approach is appropriate.

Data Sharing and Privacy

If the study involves human subject data, ensure that all data sharing complies with institutional review board requirements and applicable policies. Consult with the institutional privacy office before depositing data in public repositories.

Clinical Translation

If the goal is clinical implementation, consult with regulatory and clinical experts early in the process. Radiomics models intended for clinical use must meet regulatory requirements that differ substantially from research requirements.

Limitations and Interpretation Constraints

Radiomics has substantial promise, but its limitations must be understood by anyone using the workflow.

Sensitivity to Acquisition Variations

Radiomic features are sensitive to acquisition parameter variations. Studies that pool data from multiple scanners or protocols without harmonization will produce features that reflect technical differences instead of biological differences.

Lack of Standardization

Despite the Image Biomarker Standardisation Initiative, standardized processes of feature extraction and analysis methodology are still lacking among different research groups, which has made external validations infeasible in many cases.

Limited External Validation

The vast majority of radiomics studies are developmental and rarely provide a high level of evidence to justify the clinical translation of proposed models. Only 7.2 percent of radiomics quality scores in one meta-analysis reached 50 percent or more of the maximum.

Computational Considerations

Conventional tools rely on extensive handcrafted feature sets, which often result in redundancy and necessitate further optimization steps. Newer approaches such as SERFE aim to address this, but they require validation across multiple datasets and clinical contexts.

Interpretation Limits

Radiomic features are mathematical descriptors, not direct biological measurements. The biological meaning of specific features is often unclear, and associations between features and outcomes may not reflect causal relationships.

Frequently Asked Questions

What are radiomic features?

Radiomic features are quantitative metrics extracted from medical images that capture tissue and lesion characteristics such as heterogeneity and shape. They include first-order features describing intensity distributions, shape features describing geometry, and texture features describing spatial arrangements of intensities. These features may be used alone or combined with demographic, histologic, genomic, or proteomic data for clinical problem solving.

What is the radiomics pipeline?

The radiomics pipeline is the sequence of stages that converts medical images into predictive models. The stages are image acquisition, segmentation of the region of interest, image preprocessing, feature computation, feature selection, model building, and validation. Each stage involves specific decisions that affect the final results, and standardization across all stages is essential for reproducibility.

What is the radiomics quality score?

The radiomics quality score is a tool developed to assess the rigor of radiomics studies. It includes criteria covering image protocol quality, segmentation, preprocessing, feature selection, validation, and reporting. Version 2.0 adds criteria for fairness, explainability, quality control, and harmonization, and introduces radiomics readiness levels that reflect nine levels of incremental improvement toward clinical implementation.

How does image preprocessing affect radiomic features?

Image preprocessing substantially affects radiomic feature reproducibility and the performance of downstream classification models. Choices including voxel resampling, intensity normalization, intensity discretization, and filtering determine which features are stable across repeated acquisitions. In one study, excluding non-reproducible features improved classification performance, and the optimal preprocessing strategy varied across imaging sequences and evaluation tasks.

What is the Image Biomarker Standardisation Initiative?

The Image Biomarker Standardisation Initiative is an effort to harmonize feature definitions and preprocessing in radiomics. It specifies the mathematics of radiomic features and upstream steps including interpolation, resegmentation, intensity discretization, convolutional filtering, and feature aggregation. It provides compliance resources including digital phantoms, benchmark values, and validation checklists, and is embedded in widely used platforms such as PyRadiomics and LIFEx.

How is segmentation quality assessed in radiomics?

Segmentation quality is assessed using both visual inspection and quantitative metrics. The dice similarity coefficient, Jaccard index, and average surface distance measure agreement between automated and manual segmentations. In a study of hypopharyngeal cancer on MRI, automated segmentation models achieved dice similarity coefficients of 0.75 to 0.77, and first-order features extracted from automated segmentations showed high agreement with manual delineation.

Why is external validation important in radiomics?

External validation evaluates a model on data from a different institution, scanner, or patient population than the data used to develop the model. It is the gold standard for assessing generalizability, yet it remains rare in the radiomics literature. Models that perform well on training data but fail on external data are not clinically useful.

What are the main limitations of radiomics?

The main limitations are sensitivity to acquisition parameter variations, lack of standardized feature extraction and analysis methodology across research groups, limited external validation, and the high dimensionality of feature sets relative to typical sample sizes. These limitations impede clinical translation and must be addressed through rigorous methodology and reporting.

Related Bioinformatics Guides

• FASTQ File Format: Decoding Phred Quality Scores and Quality Control Workflows
• Cryo-EM Image Processing and 3D Reconstruction
• Docker and Containerization in Reproducible Research
• Deep Learning for Histopathology Image Classification in Veterinary Oncology
• Orchestrating Bioinformatics at Scale: Workflows, Containers, and Cloud Infrastructures

References and Further Reading

• EMBL-EBI Training. European Bioinformatics Institute.
• NCBI Data Resources. National Center for Biotechnology Information.
• Genomic Data Sharing Policy. National Institutes of Health.
• The FAIR Guiding Principles. Scientific Data.
• Introduction to Radiomics.. Journal of nuclear medicine : official publication, Society of Nuclear Medicine, 2020.
• Computational Radiomics System to Decode the Radiographic Phenotype.. Cancer research, 2017.
• Artificial intelligence-based MRI radiomics and radiogenomics in glioma.. Cancer imaging : the official publication of the International Cancer Imaging Society, 2024.
• Fully automated segmentation and radiomics feature extraction of hypopharyngeal cancer on MRI using deep learning.. European radiology, 2023.
• Radiomics in immuno-oncology.. Immuno-oncology technology, 2021.
• Spectral Entropic Radiomics Feature Extraction (SERFE): an adaptive approach for glioblastoma disease classification.. Frontiers in artificial intelligence, 2025.
• Radiomics in medical imaging-"how-to" guide and critical reflection.. Insights into imaging, 2020.
• Radiomics: A Primer on Processing Workflow and Analysis.. Seminars in ultrasound, CT, and MR, 2022.
• The long and winding road of radiomics: learnings from two meta-analyses of the radiomics quality score.. 2026.
• Radiomics Quality Score 2.0: what changed from version 1.0 and why it matters.. 2026.
• Radiomics for Predicting the Efficacy of Immunotherapy in Hepatocellular Carcinoma: A Systematic Review and Radiomics Quality Score Assessment.. 2026.
• Reply to 'Radiomics Quality Score 2.0: what changed from version 1.0 and why it matters'.. 2026.
• Correction: Evaluating the impact of the radiomics quality score: a systematic review and meta-analysis.. 2025.
• Radiomics from Routine CT and PET/CT Imaging in Laryngeal Squamous Cell Carcinoma: A Systematic Review with Radiomics Quality Score Assessment.. 2026.
• Radiomics Quality Score 2.0: towards radiomics readiness levels and clinical translation for personalized medicine.. 2025.
• Evaluating the impact of the Radiomics Quality Score: a systematic review and meta-analysis.. 2025.
• Impact of image preprocessing methods on MRI radiomics feature variability and classification performance in Parkinson’s disease motor subtype analysis. Scientific Reports, 2025.
• Correction: Impact of image preprocessing methods on MRI radiomics feature variability and classification performance in Parkinson’s disease motor subtype analysis. Scientific Reports, 2025.
• Radiomics and the Image Biomarker Standardisation Initiative (IBSI): A Narrative Review Using a Six-Question Map and Implementation Framework for Reproducible Imaging Biomarkers. Cureus, 2025.
• Influence of image preprocessing on reproducibility and longitudinal repeatability analysis of radiomics features in magnetic resonance image-guided accelerator imaging. Quantitative Imaging in Medicine and Surgery, 2026.
• Impact of image preprocessing on the volume dependence and prognostic potential of radiomics features in non-small cell lung cancer. 2016.
• Statistical Analysis on Impact of Image Preprocessing of CT Texture Patterns and Its CT Radiomic Feature Stability: A Phantom Study. Asian Pacific Journal of Cancer Prevention, 2023.
• Radiomics: A New Biomedical Workflow to Create a Predictive Model. Communications in Computer and Information Science, 2020.
• Development and optimisation of a preclinical cone beam computed tomography-based radiomics workflow for radiation oncology research. Physics and Imaging in Radiation Oncology, 2023.
• RadiomixNet: Integrating Radiomics and Feature Extraction for Advanced Pneumonia Diagnosis. IEEE Access, 2025.

This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.