Full-Field Digital Mammography (FFDM)

Unlike computed radiography (CR) which uses a cassette, FFDM employs a flat-panel detector—typically amorphous selenium (a-Se) or cesium iodide (CsI) scintillator coupled to a TFT array—to convert x-ray photons directly into an electronic signal. This eliminates the intermediate optical scanning step, resulting in higher detective quantum efficiency (DQE) and a wider dynamic range. The immediate digital readout allows for rapid image availability and post-processing before radiologist review.

FFDM acquires a single, compressed 2D projection of the breast, typically in two standard views:

• Craniocaudal (CC): A top-down projection visualizing the central, subareolar, and medial tissue.
• Mediolateral Oblique (MLO): An angled lateral projection that maximizes visualization of the upper outer quadrant and axillary tail. This fixed geometry simplifies AI model training by providing a consistent coordinate system, but introduces tissue superposition, where overlapping fibroglandular structures can obscure lesions or mimic masses.

Raw detector data undergoes a mandatory vendor-specific processing chain before display. Key algorithms include:

• Flat-field correction: Compensates for detector gain non-uniformities.
• Peripheral equalization: Brightens the skin line and subcutaneous tissue to prevent burnout.
• Contrast-limited adaptive histogram equalization (CLAHE): Enhances local contrast in dense regions without amplifying noise. AI developers must account for this processing variability across manufacturers (Hologic, GE, Siemens), as models trained on 'for-presentation' images may not generalize to 'for-processing' raw data.

FFDM systems balance spatial resolution against quantum noise. A typical detector pixel pitch of 70 µm yields a Nyquist frequency of ~7 lp/mm, sufficient for resolving fine microcalcifications (100-300 µm). However, the limiting factor is often quantum mottle at low dose levels. AI detection models must be robust to this Poisson-distributed noise, often leveraging multi-scale feature extractors that can distinguish true high-frequency calcification edges from stochastic noise spikes.

FFDM images are stored in the DICOM MG (Mammography) SOP Class, which includes critical metadata tags for AI preprocessing:

• View Position (0018,5101): CC vs. MLO for multi-view correlation logic.
• Laterality (0020,0060): Left vs. right breast.
• Organ Exposed (0018,1110): Confirms breast anatomy.
• Window Center/Width (0028,1050/1051): Default display contrast. Robust AI pipelines parse these tags to automatically orient images, apply view-specific models, and ensure correct laterality pairing before running inference.

A critical physical component of FFDM acquisition is mechanical compression, which reduces breast thickness to 4-6 cm. This serves three purposes for AI:

• Reduces scatter radiation, improving contrast.
• Minimizes geometric blur by bringing lesions closer to the detector.
• Immobilizes the patient, preventing motion artifacts. However, compression also distorts 3D anatomy into a 2D plane, creating the summation artifact—a key source of false positives where overlapping normal tissue mimics a mass. AI models must learn to differentiate true architectural distortion from pseudo-lesions caused by compression.

The algorithmic identification of tiny calcium deposits within breast tissue, which are a primary radiological sign of ductal carcinoma in situ (DCIS). These appear as bright, punctate spots on FFDM images.

• Morphology: Analyzes shape, size, and clustering patterns
• Distribution: Segmental or linear arrangements carry higher suspicion
• Contrast: High-pass filtering enhances subtle calcifications against dense parenchyma

A subtle mammographic finding characterized by radiating lines or a focal retraction of normal breast parenchyma without a visible central mass. It is often associated with invasive lobular carcinoma and represents one of the most challenging detection tasks in FFDM.

• No central density: Distinguished from spiculated masses
• Tethering patterns: Fine lines converging to a focal point
• Temporal asymmetry: Often more visible when compared to prior exams

An advanced 3D mammography technique that acquires multiple low-dose projection images over an arc to reconstruct thin slices of the breast. DBT directly addresses FFDM's primary limitation: tissue superposition.

• Slice thickness: Typically 1mm reconstructed intervals
• Projection count: 9-25 low-dose exposures over 15-50 degrees
• Clinical impact: Reduces recall rates by 15-30% compared to FFDM alone

A post-processing AI technique designed to suppress erroneous marks generated by a detection model, thereby improving specificity and reducing unnecessary recall rates. This is critical for FFDM-based CADe systems where overlapping tissue creates abundant false signals.

• Feature reclassification: Secondary model evaluates each candidate ROI
• Contextual reasoning: Considers bilateral symmetry and anatomical constraints
• Threshold optimization: Balances sensitivity against acceptable false-mark rates

An algorithmic process that geometrically links findings across the Craniocaudal (CC) and Mediolateral Oblique (MLO) views to confirm a true lesion. A finding visible in only one view is more likely a false positive caused by tissue summation.

• Epipolar geometry: Constrains the search space for matching lesions
• Nipple-to-lesion distance: Used as a coordinate reference across views
• Triangulation: Enables 3D localization from two 2D projections

The automated assignment of an ACR density category (A through D) based on the ratio of fibroglandular tissue to adipose tissue visible on FFDM. Dense tissue (categories C and D) appears white, identical to malignancies, creating a masking effect.

• Category A: Almost entirely fatty
• Category B: Scattered fibroglandular densities
• Category C: Heterogeneously dense (may obscure masses)
• Category D: Extremely dense (lowers mammographic sensitivity)