Volume of Interest (VOI)

Unlike a 2D Region of Interest (ROI) confined to a single slice, a VOI encompasses a contiguous anatomical structure across multiple axial, coronal, and sagittal planes. This volumetric contour captures the full spatial extent of a tumor or organ, enabling extraction of shape features such as sphericity, surface-to-volume ratio, and compactness that are impossible to derive from 2D analysis alone.

A VOI is composed of discrete volumetric elements called voxels, each representing a specific intensity value (e.g., Hounsfield Units in CT). The total number of voxels within the VOI directly determines the sampling resolution for subsequent texture analysis. Key considerations include:

• Isotropic resampling ensures consistent spatial measurements across all axes
• Partial volume effects at lesion boundaries can introduce intensity averaging artifacts
• Minimum VOI size thresholds are critical for statistically robust texture matrix computation

VOI delineation can be performed through several approaches, each with distinct reproducibility profiles:

• Manual segmentation: Expert clinician traces boundaries slice-by-slice; considered ground truth but suffers from high inter-observer variability
• Semi-automated: Region-growing algorithms or thresholding with manual refinement; balances efficiency and accuracy
• Fully automated: Deep learning models (e.g., nnU-Net) predict VOI boundaries without human intervention; requires rigorous validation against expert consensus
• Atlas-based: Deformable registration maps a predefined anatomical atlas onto patient-specific anatomy

A well-defined VOI serves as the computational mask for extracting multiple radiomic feature classes:

• First-order statistics: Mean, median, skewness, and kurtosis of voxel intensity distributions
• Shape features: Volume, surface area, maximum 3D diameter, and elongation
• Texture matrices: GLCM, GLRLM, GLSZM, and NGTDM computed across the entire 3D volume
• Wavelet-transformed features: Multi-scale decompositions applied to the full volumetric data

The 3D nature of VOI analysis captures cross-slice texture patterns invisible to 2D ROI methods.

VOI-based radiomic features are sensitive to multiple sources of variation that must be controlled for multicenter studies:

• Inter-observer variability: Different clinicians produce different VOI boundaries, particularly at tumor margins
• Scanner variability: Differences in reconstruction kernels and acquisition protocols across vendors
• Respiratory motion: Lesions in the thorax and upper abdomen exhibit positional shifts between slices
• Test-retest stability: Only features with high intraclass correlation coefficient (ICC) values should be retained

ComBat harmonization is commonly applied to mitigate batch effects across imaging centers.

VOI-based radiomic analysis has demonstrated utility across multiple oncological and non-oncological domains:

• Oncology: Predicting treatment response, recurrence risk, and overall survival from pretreatment CT, MRI, or PET volumes
• Cardiovascular: Quantifying atherosclerotic plaque burden and myocardial tissue characterization
• Neurology: Volumetric analysis of brain structures for neurodegenerative disease staging
• Pulmonology: Assessing interstitial lung disease extent and progression

Delta-radiomics extends VOI analysis longitudinally, tracking feature changes within the same anatomical volume across treatment timepoints.