Delta-Radiomics

The core mechanism involves calculating the net change or rate of change in radiomic features between two or more time points. Common mathematical formulations include:

• Absolute Delta: ΔF = F_time2 − F_time1
• Relative Delta: ΔF% = (F_time2 − F_time1) / F_time1 × 100 This approach captures dynamic biological processes like necrosis, fibrosis, or immune infiltration that static single-time-point imaging cannot resolve.

Delta-radiomic features have demonstrated superior predictive performance compared to baseline features alone for pathologic complete response (pCR) and progression-free survival. Key findings include:

• Changes in GLCM entropy and GLRLM run-length non-uniformity correlate with tumor heterogeneity shifts during chemotherapy
• Early delta features (2-4 weeks post-treatment initiation) can identify non-responders before anatomical shrinkage occurs
• Multi-parametric delta signatures combining CT, MRI, and PET features outperform single-modality approaches

Delta-radiomics is uniquely suited for evaluating immune checkpoint inhibitor responses, where conventional RECIST criteria often fail due to pseudoprogression. Relevant temporal patterns include:

• Transient increase in tumor volume followed by stabilization or regression
• Decreasing CT Hounsfield Unit heterogeneity indicating reduced intratumoral perfusion heterogeneity
• Shifts in PET SUVmax skewness reflecting metabolic reprogramming of the tumor microenvironment

A critical methodological requirement is establishing the threshold of meaningful change beyond measurement noise. This involves:

• Calculating the repeatability coefficient (RC) from test-retest cohorts scanned within a short interval without intervention
• Defining the 95% confidence interval for the delta of each feature
• Only features with a delta exceeding the RC are considered true biological change
• Concordance correlation coefficient (CCC) values > 0.85 are typically required for feature inclusion in delta models

Accurate delta calculation depends on precise spatial alignment of longitudinal scans. Essential preprocessing steps include:

• Rigid or deformable registration to align baseline and follow-up images to a common coordinate space
• Voxel resampling to ensure isotropic resolution across time points
• Intensity normalization using histogram matching or ComBat harmonization to correct for scanner drift
• Failure to register can introduce spurious delta values from anatomical misalignment rather than biological change

Published delta-radiomic signatures have been validated for multiple clinical endpoints:

• Non-small cell lung cancer: Delta-GLCM contrast and delta-GLSZM zone percentage predict overall survival after chemoradiation
• Glioblastoma: Delta-wavelet features differentiate true progression from pseudoprogression with AUC > 0.85
• Rectal cancer: Delta-shape sphericity and delta-first-order kurtosis predict pathologic complete response to neoadjuvant therapy These signatures are typically combined using Cox proportional hazards or logistic regression models.

The high-throughput extraction and analysis of quantitative features from medical images to create mineable data for decision support. Radiomics converts standard-of-care CT, MRI, and PET scans into high-dimensional feature spaces capturing tumor intensity, shape, and texture. Delta-radiomics extends this paradigm by applying feature extraction to longitudinal image pairs, enabling the quantification of temporal change rather than static snapshots.

A set of mathematical methods for quantifying the spatial arrangement of pixel or voxel intensities to characterize tissue heterogeneity. Key matrices include:

• GLCM: Captures second-order relationships between pixel pairs
• GLRLM: Quantifies runs of consecutive same-value pixels
• GLSZM: Measures connected regions of identical intensity In delta-radiomics, changes in texture heterogeneity often precede volumetric changes, serving as early indicators of treatment response.

The computational process of removing unwanted technical variability from radiomic features caused by differences in scanner manufacturers, acquisition protocols, or reconstruction algorithms. ComBat harmonization, adapted from genomics, is the most widely used method for multi-center delta-radiomics studies. Without harmonization, scanner-induced variation can mask or mimic genuine biological change, rendering longitudinal comparisons unreliable.

A composite biomarker consisting of a specific set of weighted radiomic features combined via a mathematical model to predict a clinical endpoint. Delta-radiomic signatures incorporate temporal feature deltas—the arithmetic or relative change in individual features between timepoints—as input variables. These signatures are typically developed using LASSO or elastic net regression to handle the high-dimensional feature space while avoiding overfitting.

The use of deep convolutional neural networks to automatically learn hierarchical feature representations directly from medical images, bypassing handcrafted feature engineering. In the delta context, Siamese network architectures process paired pre- and post-treatment scans through shared weights, learning to represent temporal change in a latent feature space. This approach can capture non-linear temporal dynamics that predefined mathematical features may miss.

The assessment of the stability of radiomic feature measurements when imaging is repeated on the same subject under identical conditions within a short interval. Features with poor test-retest reliability are excluded from delta-radiomics models, as their temporal variation may reflect measurement noise rather than biological change. Intraclass correlation coefficient (ICC) values above 0.75 are typically required for a feature to be considered reproducible enough for longitudinal analysis.