Radiomics: Feature Extraction and Analysis

Module: Healthcare AI | Difficulty: Advanced

GLCM Energy

GLCM Entropy

Wavelet Features

Radiomics Feature Categories

| Category | Feature Count | Example Features | |----------|--------------|------------------| | First-order | 18 | Mean, variance, skewness, kurtosis | | Shape | 14 | Volume, surface area, sphericity | | GLCM | 24 | Contrast, correlation, energy | | GLRLM | 16 | SRE, LRE, GLNU | | GLSZM | 16 | SAE, LAE, ZSN | | NGTDM | 5 | Coarseness, contrast, busyness |

import numpy as np
from skimage.feature import graycomatrix, graycoprops

def extract_first_order_features(volume):
    features = {}
    features['mean'] = np.mean(volume)
    features['std'] = np.std(volume)
    features['skewness'] = float(np.mean(
        ((volume - features['mean']) / (features['std'] + 1e-8))**3))
    features['kurtosis'] = float(np.mean(
        ((volume - features['mean']) / (features['std'] + 1e-8))**4) - 3)
    features['energy'] = float(np.sum(volume**2))
    return features

def extract_glcm_features(volume_2d, distances=[1, 2, 4], angles=[0, np.pi/4, np.pi/2]):
    volume_2d = (volume_2d / volume_2d.max() * 255).astype(np.uint8)
    glcm = graycomatrix(volume_2d, distances=distances, angles=angles,
                        levels=256, symmetric=True, normed=True)
    features = {}
    for prop in ['contrast', 'dissimilarity', 'homogeneity', 'energy', 'correlation']:
        vals = graycoprops(glcm, prop)
        features[f'glcm_{prop}_mean'] = float(np.mean(vals))
        features[f'glcm_{prop}_std'] = float(np.std(vals))
    return features

def extract_wavelet_features(volume, wavelet='db4'):
    import pywt
    coeffs = pywt.wavedec3(volume, wavelet, level=2)
    features = {}
    for i, coeff in enumerate(coeffs):
        if isinstance(coeff, np.ndarray):
            features[f'wavelet_level{i}_mean'] = float(np.mean(np.abs(coeff)))
            features[f'wavelet_level{i}_energy'] = float(np.sum(coeff**2))
    return features

def build_radiomics_pipeline(volume, mask):
    masked = volume * mask
    features = {}
    features.update(extract_first_order_features(masked[mask > 0]))
    slice_idx = volume.shape[2] // 2
    features.update(extract_glcm_features(masked[:, :, slice_idx]))
    return features

volume = np.random.rand(64, 64, 64)
mask = (volume > 0.3).astype(float)
features = build_radiomics_pipeline(volume, mask)
print(f'Extracted {len(features)} radiomics features')

Research Insight: Radiomics features are susceptible to scan protocol variations. The Image Biomarker Standardization Initiative (IBSI) provides standardized feature definitions and phantom-based validation. ComBat harmonization and other batch effect correction methods are essential when combining data from multiple institutions.