Decoding Tissue Heterogeneity Through Quantitative Radiomics

256 concepts232 papers9 proposals

When a radiologist reviews a medical image, they interpret it through a qualitative lens, identifying macroscopic shapes, distinct boundaries, and gross density variations. However, a digital medical image—whether a CT, MRI, or PET scan—is fundamentally a three-dimensional grid of voxels containing precise numerical data. Hidden within these voxel distributions are subtle, sub-visual patterns of spatial variation and texture that reflect the underlying microstructural and biological properties of the tissue.

The field of radiomics bridges this gap by converting standard medical images into high-dimensional, mineable quantitative data. As shown by Hugo J. W. L. Aerts et al. in Corrigendum: Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach, radiomics extracts a massive array of quantitative features to infer complex tissue phenotypes non-invasively. This approach acts as a "virtual biopsy," mapping macroscopic imaging textures to microscopic cellular and genetic characteristics.

The Hierarchy of Radiomic Features

To understand why radiomics works, we must look at how it mathematically formalizes visual and sub-visual patterns. Feature extraction typically follows a structured hierarchy, moving from simple global statistics to complex spatial relationships.

First-Order Statistics

These features describe the distribution of individual voxel intensities within a region of interest (ROI), completely ignoring their spatial arrangement. Metrics like the mean or median provide basic density information, while higher-order moments like skewness (the asymmetry of the distribution) and kurtosis (the "peakedness" of the distribution) capture subtle shifts in tissue composition. For example, Song Xue et al. demonstrated in Application of machine learning to pretherapeutically estimate dosimetry in men with advanced prostate cancer treated with 177Lu-PSMA I&T therapy that voxel-level intensity statistics can capture spatial heterogeneity in organ uptake that standard population-based averages ignore.

Shape and Morphological Descriptors

These features quantify the geometric properties of a segmented region. Metrics include volume, surface-to-volume ratio, sphericity, and flatness. Peng Jiang et al. leveraged automated shape analysis in Development and validation of an institutional nomogram for aiding aneurysm rupture risk stratification to objectively evaluate anatomical irregularities, providing a more reliable risk assessment than subjective human observation.

Texture and Higher-Order Spatial Features

This is where radiomics unlocks patterns invisible to the human eye. Texture analysis computes the spatial relationships between neighboring voxels using mathematical constructs like Gray Level Co-occurrence Matrices (GLCM) or Gray Level Run Length Matrices (GLRLM). These matrices calculate how often pairs of voxels with specific intensities occur in specific spatial arrangements, yielding metrics like entropy (a measure of randomness or heterogeneity) and contrast.

Zhenyu Shu utilized these sub-visual texture matrices in A novel computed tomography radiomic nomogram for early evaluation of small airway dysfunction development to detect microstructural lung changes long before they manifest as visible clinical disease.

[Raw Medical Image] 
       │
       ▼
[Segmentation of ROI] 
       │
       ├─► First-Order Stats (Voxel Intensity Histograms: Skewness, Kurtosis)
       ├─► Shape Descriptors (Sphericity, Surface-to-Volume Ratio)
       └─► Texture Matrices  (Spatial Relationships: GLCM Entropy, Contrast)

Transforming Images to Expose Hidden Scales

Raw voxel intensities are often subject to noise and scanner variations. To isolate specific spatial frequencies and directional patterns, radiomics applies mathematical filters to the images before feature extraction.

Common transformations include Wavelet transforms and Laplacian of Gaussian (LoG) filters. Wavelet transforms decompose the image into high- and low-frequency components along different spatial dimensions, effectively isolating fine textures from coarse structural backgrounds. LoG filters highlight edges and areas of rapid intensity change at configurable scales.

As shown by Yuxue Xie et al. in Radiomics Feature Analysis of Cartilage and Subchondral Bone in Differentiating Knees Predisposed to Posttraumatic Osteoarthritis after Anterior Cruciate Ligament Reconstruction from Healthy Knees, extracting features from these transformed spaces provides a far more comprehensive and sensitive representation of tissue health than traditional single-value clinical metrics.

The fundamental premise of radiomics is that macroscopic imaging patterns are a direct manifestation of microscopic biological processes. A tumor, for instance, is not a uniform mass of identical cells; it is a highly heterogeneous environment composed of varying cellular densities, necrotic cores, localized hypoxia, and complex vascular networks.

When these microscopic variations occur, they alter the local physical properties of the tissue—such as X-ray attenuation in CT or proton relaxation times in MRI. This creates subtle, localized fluctuations in voxel intensities. While a clinician cannot track the statistical variance of thousands of pixels across a 3D volume, a machine learning model trained on radiomic features can.

For example, Rui Gu et al. demonstrated in CT radiomics prediction of CXCL9 expression and survival in ovarian cancer that macroscopic gray-level distributions can accurately predict microscopic protein expression and tumor immune status. Similarly, Saima Rathore et al. in Multivariate Analysis of Preoperative Magnetic Resonance Imaging Reveals Transcriptomic Classification of de novo Glioblastoma Patients showed that multi-parametric MRI textures map directly onto distinct transcriptomic subtypes of glioblastoma, bypassing the immediate need for invasive tissue biopsies.

Radiomics vs. Deep Learning: The Power of Inductive Biases

With the rise of deep learning, one might wonder why hand-crafted radiomic features remain vital instead of simply feeding raw 3D images into convolutional neural networks or transformers. The answer lies in data efficiency, interpretability, and noise filtering.

Deep learning models require massive datasets to learn robust representations from scratch. In medical imaging, datasets are frequently small and highly heterogeneous due to variations in scanner manufacturers and imaging protocols. Hand-crafted radiomic features act as a powerful form of feature engineering, embedding strong mathematical inductive biases (like shape, contrast, and spatial frequency) directly into the pipeline.

Dmitrii Seletkov et al. highlighted this advantage in AI-driven preclinical disease risk assessment using imaging in UK biobank, finding that structured radiomic features effectively filter out imaging noise and artifacts that can otherwise destabilize deep learning models trained on raw 3D volumes. Furthermore, as noted by Sobia Yousaf in Brain Tumor Survival Prediction using Radiomics Features, engineered features like Haralick descriptors capture tissue heterogeneity explicitly. This allows simpler, highly interpretable classifiers like Random Forests or Support Vector Machines to match or exceed deep learning performance in data-constrained clinical environments.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_radiomic_feature_extraction_0a59c8fd,
  title = {radiomic feature extraction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/radiomic-feature-extraction/rdr_0a59c8fde82144719c171335c4aec75c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_radiomic_feature_extraction_0a59c8fd_v1,
  title = {radiomic feature extraction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/radiomic-feature-extraction/rdr_0a59c8fde82144719c171335c4aec75c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake