Radiomics and Medical Imaging AI

Under scientific review

This page describes quantitative imaging features and machine learning workflows in radiology and pathology. Use it for orientation, not clinical decision-making.

Medical imaging scans (CT, MRI, PET, Mammography) are not merely pictures to be viewed by eye — they are high-dimensional quantitative spatial datasets. Radiomics is the automated extraction of high-throughput quantitative features from medical images to convert visual phenotypes into mineable data.

For computer vision engineers, radiomics bridges traditional handcrafted image processing (texture matrices, shape descriptors) and modern deep learning embeddings (CNNs, Vision Transformers) to predict tumor histology, driver mutations (Radiogenomics), and therapy response.

---

1. The Radiomics Processing Pipeline

[ DICOM Scans (CT/MRI/PET) ] ──► [ Preprocessing & Resampling ] ──► [ ROI Segmentation (3D Mask) ]
                                                                             │
[ Clinical / Genomic Outcome ] ◄── [ ML Model (ElasticNet/XGBoost) ] ◄── [ Feature Extraction (PyRadiomics) ]

1. Acquisition & Metadata Parsing: DICOM images are ingested along with acquisition parameters (slice thickness, kilovoltage, contrast timing, reconstruction kernels).
2. Preprocessing & Standardization:
   • Voxel Resampling: Resampling images to isotropic voxel spacing (e.g., $1\text{mm} \times 1\text{mm} \times 1\text{mm}$) using B-spline interpolation.
   • Intensity Normalization: Z-score or min-max normalization, especially critical for MRI where signal intensity units are arbitrary (unlike CT's calibrated Hounsfield Units - HU).
3. Segmentation (Mask Creation): Defining the 3D Region of Interest (ROI) for the primary tumor and sub-regions (necrotic core, enhancing rim, peritumoral edema). Done manually by radiologists or automatically via deep learning models (nnU-Net).
4. Feature Extraction: Extracting mathematical descriptors conforming to the Image Biomarker Standardisation Initiative (IBSI).
5. Machine Learning & Modeling: Feature selection (mRMR, Lasso) followed by model training, rigorous out-of-distribution (OOD) testing, and probability calibration.

---

2. Feature Families (PyRadiomics Standard)

Radiomics extracts hundreds of features categorized into mathematical families:

Feature Family	What It Measures	Computer Vision Analogy
Shape & Size	Volume, surface area, sphericity, compactness, elongation	3D mesh geometry & bounding box ratios
First-Order (Intensity)	Mean, median, standard deviation, skewness, kurtosis, entropy	Grayscale intensity histogram statistics
GLCM (Gray Level Co-occurrence Matrix)	Contrast, correlation, energy, homogeneity	Spatial arrangement of pixel pair intensities
GLRLM / GLSZM (Run/Zone Length)	Coarseness, texture homogeneity, regional micro-patterns	Spatial runs of consecutive identical pixel values
Filter-Derived Features	High/low spatial frequencies extracted via Wavelet or Laplacian of Gaussian (LoG) filters	Multi-scale edge & blob detection (DoG/Sobel filters)

---

3. Radiogenomics: Predicting Mutations from Scans

Radiogenomics correlates radiomic texture features with underlying genomic alterations, aiming to create non-invasive "virtual biopsies":

• Brain (Glioma): Predicting IDH1 mutation status or 1p/19q co-deletion from MRI fluid-attenuated inversion recovery (FLAIR) signal heterogeneity.
• Lung (NSCLC): Predicting EGFR mutation or ALK fusion status from CT spiculated margins and ground-glass opacity patterns.
• Breast: Correlating mammographic density and DCE-MRI enhancement kinetics with BRCA1/2 status or Oncotype DX recurrence scores.

---

4. Key Engineering Pitfalls in Radiomics

• Inter-Scanner Variability: Features extracted from a Siemens CT scanner may fail on a GE scanner due to hardware detector differences and reconstruction kernel variations. Solutions include ComBat harmonization applied directly to feature matrices.
• Segmentation Sensitivity: Features like sphericity or GLCM contrast are highly sensitive to small manual segmentation boundary shifts. Features must be audited for inter-observer agreement (Intraclass Correlation Coefficient - ICC > 0.85).

---

References

1. Zwanenburg, A., et al. (2020). The Image Biomarker Standardisation Initiative: standardized quantitative radiomics for high-throughput image-based phenotyping. Radiology, 295(2), 328-338.
2. van Griethuysen, J. J., et al. (2017). Computational radiomics system to decode the radiographic phenotype. Cancer Research, 77(21), e104-e107.
3. Gillies, R. J., Kinahan, P. E., & Hricak, H. (2016). Radiomics: images are more than pictures, they are data. Radiology, 278(2), 563-577.