Breast Imaging Reporting and Data System (BI-RADS)

The core of BI-RADS is a numerical scale from 0 to 6 that encodes the radiologist's level of suspicion and recommended action. Each category maps to a specific management pathway, eliminating ambiguous prose from reports.

• Category 0 (Incomplete): Requires additional imaging or prior exams for comparison.
• Category 1 (Negative): Routine screening recommended; 0% likelihood of malignancy.
• Category 2 (Benign): Definitively benign finding (e.g., calcified fibroadenoma); routine screening.
• Category 3 (Probably Benign): < 2% risk of malignancy; short-interval follow-up recommended.
• Category 4 (Suspicious): Biopsy recommended; subdivided into 4A (low), 4B (moderate), and 4C (high suspicion).
• Category 5 (Highly Suggestive): ≥ 95% risk of malignancy; biopsy and surgical planning required.
• Category 6 (Known Biopsy-Proven): Malignancy confirmed via biopsy; used for staging and treatment response.

BI-RADS mandates precise terminology to describe mass margins, shape, and density, transforming subjective visual impressions into quantifiable features. This structured lexicon is critical for training AI models on consistent ground-truth labels.

• Margin descriptors are the most predictive features: circumscribed (well-defined, likely benign), microlobulated (subtle undulations, suspicious), indistinct (ill-defined, suspicious), and spiculated (radiating lines, highly malignant).
• Shape is categorized as oval, round, or irregular; irregular shapes carry higher suspicion.
• Density relative to fibroglandular tissue: high-density masses are suspicious, while low-density or fat-containing masses are typically benign.

BI-RADS provides a detailed taxonomy for calcification morphology and distribution, which is essential for distinguishing benign from malignant microcalcifications. This granularity directly informs AI detection algorithm design.

• Typically benign morphologies: Skin (lucent-centered), vascular (tram-track), coarse or popcorn-like (involuting fibroadenoma), rod-like (secretory disease), and rim (fat necrosis).
• Suspicious morphologies: Amorphous (indistinct, hazy), coarse heterogeneous (irregular, > 0.5 mm), fine pleomorphic (varying shapes, < 0.5 mm), and fine linear/fine-linear branching (casting-type, highly associated with DCIS).
• Distribution modifies risk: diffuse or scattered calcifications are often benign, while segmental (ductal territory) or linear distributions raise suspicion significantly.

BI-RADS includes a standardized breast composition assessment that quantifies the ratio of fibroglandular tissue to fat. This is now a legal reporting requirement in many jurisdictions due to its impact on cancer masking and risk.

• ACR A (Almost Entirely Fatty): Minimal fibroglandular tissue; highest sensitivity for mammography.
• ACR B (Scattered Fibroglandular Densities): Scattered areas of density; may obscure small lesions.
• ACR C (Heterogeneously Dense): More than half of the breast is dense; may obscure small masses; sensitivity is reduced.
• ACR D (Extremely Dense): Nearly all fibroglandular tissue; significantly lowers mammographic sensitivity; supplemental screening (ultrasound, MRI) is often recommended.

BI-RADS formally defines architectural distortion as a focal disruption of the normal breast parenchymal pattern without a visible mass. It is a subtle but critical finding, often the earliest mammographic sign of invasive lobular carcinoma.

• Characterized by radiating lines or spiculations emanating from a central point, with focal retraction or tethering of surrounding tissue.
• On digital breast tomosynthesis (DBT), architectural distortion is significantly more conspicuous than on 2D FFDM, making it a primary target for AI detection algorithms.
• When identified, it is typically classified as BI-RADS 4 (Suspicious) and warrants biopsy, even in the absence of a sonographic correlate.

BI-RADS extends beyond masses and calcifications to include associated features that modify suspicion and special case findings with pathognomonic appearances. These categories ensure comprehensive reporting.

• Associated features: Skin retraction, nipple retraction, skin thickening, trabecular thickening, axillary adenopathy, and architectural distortion. Their presence upgrades suspicion.
• Special cases: Includes intramammary lymph nodes (reniform shape with fatty hilum, benign), global asymmetry (seen on only one view, often normal tissue), and focal asymmetry (seen on two views, lacks convex borders of a mass).
• Developing asymmetry is a new, enlarging, or more conspicuous asymmetry compared to prior exams and is considered suspicious (BI-RADS 4) until proven otherwise.

An AI system that goes beyond marking suspicious regions to characterize a detected lesion, providing an automated assessment of disease likelihood or a specific BI-RADS category. Unlike detection-only systems, CADx algorithms analyze morphological features—margin characteristics, shape, density, and internal echo patterns—to estimate malignancy probability. Modern deep learning CADx systems are trained to map imaging features directly to BI-RADS descriptors, functioning as a second reader that offers a diagnostic suggestion rather than just a location marker.

A post-processing AI technique designed to suppress erroneous marks generated by a detection model, directly improving the positive predictive value of a BI-RADS 0 or BI-RADS 4 assessment. By analyzing false positive patterns—such as vascular calcifications, skin folds, or overlapping fibroglandular tissue—these algorithms increase specificity without sacrificing sensitivity. Effective false positive reduction is critical for maintaining low recall rates and preventing unnecessary diagnostic workups triggered by over-cautious CADe marks.

The automated assignment of ACR density categories A through D based on the ratio of fibroglandular tissue to adipose tissue in a mammogram. Dense tissue appears white on mammography—the same radiographic density as malignancy—creating a masking effect that reduces sensitivity. BI-RADS includes a mandatory density descriptor because higher density (Categories C and D) directly impacts the decision to recommend supplemental screening with ultrasound or MRI. AI-driven density assessment provides consistent, quantitative measurements, reducing inter-reader variability.

The process of adjusting a diagnostic model's output probabilities so that a predicted confidence score of 0.8 truly reflects an 80% empirical likelihood of malignancy. In BI-RADS terms, a well-calibrated model ensures that lesions assigned a high probability of malignancy actually correspond to BI-RADS 5 (highly suggestive) findings, while low-probability outputs align with BI-RADS 2 (benign). Calibration is measured using expected calibration error (ECE) and reliability diagrams, and is essential for clinical trust—an overconfident model risks unnecessary biopsies, while an underconfident one misses cancers.

An algorithmic process that geometrically links findings across the Craniocaudal (CC) and Mediolateral Oblique (MLO) views to confirm a true lesion. A true mass appears in both views at corresponding locations along the arc of compression, while superimposed tissue artifacts appear in only one. By correlating detection marks across views, AI systems can suppress false positives and increase the confidence of a BI-RADS assessment. This spatial reasoning mimics the radiologist's triangulation process and is a key component of modern deep learning detection pipelines.

The temporal evaluation of contrast agent uptake and washout patterns in Contrast-Enhanced Mammography (CEM) or breast MRI. Malignant lesions typically exhibit rapid initial enhancement followed by a washout pattern (Type III curve) due to arteriovenous shunting and increased vascular permeability. Benign lesions more often show persistent enhancement (Type I). This kinetic data provides functional information that complements the morphological BI-RADS descriptors, helping to upgrade or downgrade a BI-RADS 4 assessment into a more definitive category.