Region of Interest (ROI)

An ROI can be defined as a two-dimensional (2D) contour on a single slice or a three-dimensional (3D) volume spanning multiple slices. 3D ROIs capture the full tumoral heterogeneity and are mandatory for extracting shape features like sphericity and compactness. A 2D ROI, often used for speed, limits analysis to a single representative plane and ignores volumetric textural variations.

The method of contouring critically impacts feature robustness. Manual segmentation by an expert radiologist is the clinical gold standard but suffers from high inter-observer variability. Semi-automated methods (e.g., region growing, graph cuts) require a user-defined seed point. Fully automated deep learning models (e.g., U-Net) offer perfect reproducibility but must be rigorously validated against manual contours to avoid systematic bias.

A precise ROI definition must specify the included tissue compartment. Options include the gross tumor volume (GTV) , the clinical target volume (CTV) with microscopic spread margins, or a specific habitat sub-region like a necrotic core or enhancing rim. Mixing compartments within a single ROI introduces biological noise that dilutes the predictive power of first-order and textural features.

The treatment of the tumor boundary is a critical definition parameter. An ROI can be drawn tightly along the visible lesion edge or include a fixed dilation margin (e.g., 2mm) to capture the peri-tumoral microenvironment. Shrinking the contour inward by a small distance is sometimes used to avoid partial volume effects at the tissue interface, ensuring that voxels are purely from the target tissue class.

When an ROI is defined on one sequence (e.g., a T1-weighted MRI) and applied to another (e.g., an ADC map), rigid or deformable co-registration is required. Misalignment due to patient motion or organ deformation can cause the ROI to sample the wrong tissue in the target sequence. The Dice Similarity Coefficient is the standard metric used to validate the geometric overlap after registration.

ROIs can be refined by applying absolute intensity thresholds to exclude non-tissue voxels like air, bone, or fluid. For example, a Hounsfield Unit (HU) window can isolate soft tissue in a CT scan. This step ensures that first-order statistics (mean, minimum, maximum) are calculated only on clinically relevant voxels and are not skewed by extreme values from adjacent anatomical structures.

The critical preprocessing step of binning continuous voxel intensity values within an ROI into a finite number of discrete gray levels. This step directly determines the size and statistical properties of texture matrices.

• Fixed bin number discretization divides the intensity range into a constant number of bins regardless of ROI dynamic range
• Fixed bin width discretization uses a constant bin size in Hounsfield Units or SUV, preserving physical meaning
• Discretization choice significantly impacts GLCM, GLRLM, and GLSZM feature reproducibility

A composite biomarker consisting of a selected panel of multiple quantitative features extracted from one or more ROIs, combined via a mathematical model to predict a specific clinical endpoint.

• Built using feature selection techniques like LASSO or mRMR to avoid overfitting
• Validated using Intraclass Correlation Coefficient (ICC) to ensure ROI stability across raters
• Represents the translational endpoint of ROI analysis—transforming contour geometry into clinically actionable prognostic scores

A technique that partitions a single ROI into multiple distinct sub-regions based on voxel-wise clustering of functional or structural imaging parameters. This reveals intratumoral heterogeneity invisible to whole-ROI analysis.

• Clusters voxels using multi-parametric MRI or PET/CT data
• Each habitat sub-ROI represents a biologically distinct tumor microenvironment
• Enables spatial heterogeneity quantification beyond single-value ROI summaries