High-Dimensional Radiomics for Non-Invasive Tissue Phenotyping

21 concepts21 papers1 proposal

When a clinician evaluates an MRI, CT, or PET scan, they typically rely on macro-level visual criteria: Is the tumor boundary irregular? How large is the mass? Is the signal intensity abnormally high? While this qualitative approach is the cornerstone of modern diagnostics, it discards a vast reservoir of digital information. A medical image is not merely a picture to be looked at; it is a structured grid of quantitative data.

Radiomics is the practice of converting these digital images into high-dimensional, mineable datasets. By extracting thousands of quantitative descriptors—spanning shape, intensity distributions, and complex spatial textures—radiomics uncovers sub-visual patterns of tissue heterogeneity. This approach allows clinicians to decode the underlying biological properties of a tissue non-invasively, effectively acting as a digital or "virtual" biopsy.

Beyond the Human Eye: Images as Data

The core insight behind radiomics is that physiological and molecular variations within tissue alter its micro-architecture, which in turn alters the spatial distribution of pixel or voxel intensities in medical imaging. While human vision is optimized for recognizing macro-structures and distinct borders, it is poorly equipped to quantify subtle statistical variations across thousands of pixels.

As demonstrated by Ingrisch et al., these sub-visual patterns contain critical prognostic signatures. By analyzing standard baseline MRIs of glioblastoma patients, the authors successfully stratified individuals into distinct risk groups that were otherwise indistinguishable through standard clinical inspection. Similarly, Jiang et al. applied radiomic feature extraction to CT scans to identify a "digital signature" capable of distinguishing standard pneumonia from its severe, life-threatening forms.

By treating the image as an objective, structured dataset, radiomics removes the subjectivity of manual visual assessment and uncovers tissue characteristics that are imperceptible to human observers.

The Radiomic Pipeline: From Pixels to Features

To understand why radiomics works, it is helpful to look at how an image is systematically decomposed into quantitative features. The pipeline generally follows three core phases:

[ Medical Image ] ──> [ Segmentation (ROI) ] ──> [ Feature Extraction ] ──> [ Machine Learning Model ]

1. Segmentation

First, a Region of Interest (ROI) or Volume of Interest (VOI) is defined around the target tissue or tumor. While this can be done manually, modern workflows increasingly leverage automated deep learning frameworks. For instance, Cao et al. utilized the nnU-Net architecture to automatically segment gastric tumors from preoperative CT scans, ensuring highly reproducible boundaries before feature extraction.

2. Feature Extraction

Once the region is isolated, computational algorithms calculate hundreds of mathematical descriptors. These features generally fall into three tiers of complexity:

  • First-Order Statistics: These describe the distribution of individual voxel intensities within the region, ignoring spatial relationships. Metrics include the mean, variance, skewness, and kurtosis of the image histogram.
  • Shape and Morphological Descriptors: These quantify the 3D geometric properties of the region, such as volume, sphericity, and surface-to-volume ratio. In a multicenter study on renal tumors, single author Yaohai Wu demonstrated that geometric features—specifically a tumor's flatness—can serve as primary mathematical predictors of malignancy.
  • Higher-Order Texture Features: These measure the spatial relationships between neighboring pixels to quantify tissue heterogeneity. Algorithms like Gray-Level Co-occurrence Matrices (GLCM) calculate how often pairs of pixels with specific intensities occur next to each other in a given direction. As noted by Colson et al. and Zhang et al., these texture metrics are highly sensitive to spatial variations in tracer uptake or tissue density, capturing complex biological patterns that human eyes cannot resolve.

Why It Works: Mapping Texture to Biology

Why should the mathematical texture of pixels on a screen correlate with a patient's survival or a tumor's genetic profile? The answer lies in the relationship between micro-scale biology and macro-scale imaging physics.

Tumors are not uniform clumps of identical cells; they are highly chaotic microenvironments characterized by localized areas of rapid cell proliferation, variations in vascular density, and pockets of tissue death (necrosis).

  • An area with dense cellular packing restricts the diffusion of water molecules, altering the signal intensity on an MRI.
  • A region with leaky, chaotic blood vessels causes uneven distribution of contrast agents or radiotracers, creating high-frequency fluctuations in a CT or PET scan.

By formalizing these pixel-to-pixel relationships, radiomics maps macro-imaging textures directly back to these micro-biological variations. For example, Li et al. showed that high-dimensional metabolic features extracted from PET imaging could accurately predict the Isocitrate Dehydrogenase (IDH) genotype in glioma patients without requiring a physical tissue sample. In another application, Kaissis et al. demonstrated that these sub-visual phenotypic patterns could identify aggressive molecular subtypes of pancreatic cancer, allowing clinicians to predict which patients would respond better to specific chemotherapy regimens.

Decoding the High-Dimensional Space

Because the extraction process generates thousands of features from a single region, the resulting dataset is highly dimensional. To prevent overfitting and isolate the true diagnostic signals from background noise, machine learning algorithms are deployed to select the most predictive features and build classification models (Tabassum et al.). Algorithms such as Random Forests, XGBoost, and Logistic Regression are commonly trained on these feature vectors to map the mathematical signatures to clinical endpoints (Feng et al.).

Historically, a major barrier to adopting these models in clinical practice has been the "black box" problem—the difficulty of understanding why a complex machine learning model makes a specific prediction. To transition radiomics from an experimental software tool into a trusted medical procedure (a "virtual biopsy"), researchers have integrated explainable AI frameworks (Murray et al.; Zhu et al.).

By applying SHAP (Shapley Additive exPlanations) values to these models, authors like Liu et al. and Yaohai Wu have successfully traced automated diagnostic decisions back to the exact morphological or textural features that drove the prediction. This transparency ensures that the mathematical models remain grounded in interpretable tissue physiology, paving the way for objective, non-invasive clinical decision support.

Key Foundational Work

Radiomic Analysis Reveals Prognostic Information in T1-Weighted Baseline Magnetic Resonance Imaging in Patients With Glioblastoma (2017) — contributed 1 concept elements to this direction.
Dermal Radiomics for Melanoma Screening (2015) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_quantitative_image_phenotyping_1a03768e,
  title = {quantitative image phenotyping},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/quantitative-image-phenotyping/rdr_1a03768ecd00476cb6ae04f2d81b2d24},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_quantitative_image_phenotyping_1a03768e_v1,
  title = {quantitative image phenotyping},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/quantitative-image-phenotyping/rdr_1a03768ecd00476cb6ae04f2d81b2d24/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake