Region of Interest (ROI)

A valid ROI must precisely delineate the suspect tissue, distinguishing it from background parenchyma. This involves defining a bounding box for detection tasks or a segmentation mask for pixel-level analysis. Accurate localization is critical for correlating findings across multi-view correlation (CC and MLO views) and for guiding biopsy. The centroid of the ROI is often used for prior exam registration to enable temporal comparison.

The ROI encapsulates critical morphological signatures that distinguish benign from malignant findings. Key features include:

• Spiculation: Radiating lines from a mass margin, a highly specific malignancy indicator.
• Margin characteristics: Circumscribed vs. indistinct boundaries.
• Shape: Round, oval, or irregular.
• Internal density: Homogeneous vs. heterogeneous. These features are the basis for BI-RADS lexicon descriptors and are learned by deep patch-based analysis models.

Every ROI must be associated with a confidence score reflecting the algorithm's certainty that it represents a true lesion. This score enables worklist prioritization, where high-suspicion cases are triaged to the top of the reading queue. Model calibration ensures this score is a statistically reliable estimate of malignancy likelihood, not just a ranking metric. A well-calibrated model's 80% confidence should correspond to an 80% empirical true positive rate.

An isolated ROI has limited value. A well-defined ROI is linked to its anatomical context and historical priors. This includes:

• Multi-view correlation: Linking a finding in the CC view to its counterpart in the MLO view to confirm it is a real 3D structure.
• Temporal comparison: Registering the ROI against a prior mammogram to detect interval change, such as new calcifications or architectural distortion.
• Breast density context: Weighting suspicion based on the masking effect of dense fibroglandular tissue.

A robust ROI definition includes mechanisms to suppress non-lesion signals. False positive reduction algorithms analyze candidate ROIs to reject artifacts such as:

• Skin folds and vascular calcifications.
• Motion blur or detector artifacts.
• Superimposed tissue mimicking a mass in 2D FFDM. This is evaluated using Free-Response Operating Characteristic (FROC) curves, which plot sensitivity against the average number of false positives per image.

Beyond visual detection, a precisely segmented ROI serves as a substrate for radiomics feature extraction. This involves mining high-throughput quantitative features—such as texture, entropy, and wavelet coefficients—from the delineated pixels. These features can be fed into downstream Computer-Aided Diagnosis (CADx) models to predict tumor grade, receptor status, or treatment response, moving the ROI from a localization tool to a prognostic biomarker.