Digital Breast Tomosynthesis (DBT)

The X-ray tube moves in a limited arc (typically 15° to 50°) over the compressed breast, acquiring 9 to 25 low-dose projection images. Unlike CT, which uses a full 360° rotation, DBT employs an incomplete angular sampling to minimize radiation dose while providing depth resolution. The resulting projection dataset is reconstructed using iterative algorithms such as Filtered Back Projection (FBP) or Simultaneous Algebraic Reconstruction Technique (SART) to generate a stack of 1mm-thick slices parallel to the detector plane.

The primary clinical advantage of DBT is the reduction of anatomical noise caused by overlapping fibroglandular tissue. In 2D Full-Field Digital Mammography (FFDM), a malignancy can be obscured by normal parenchyma above or below it. DBT resolves this by isolating structures into discrete slices, allowing radiologists to scroll through the breast volume. This mechanism directly addresses the masking effect prevalent in dense breasts (ACR categories C and D), where sensitivity in FFDM can drop to 30-48%.

To reduce total radiation dose, modern DBT systems generate a Synthesized 2D (s2D) image from the tomographic volume, eliminating the need for a separate FFDM acquisition. Algorithms such as Maximum Intensity Projection (MIP) or Volume-Weighted Projection condense the 3D data into a familiar 2D representation. The FDA-approved combo-mode (DBT + s2D) delivers a dose comparable to a standard 2D mammogram while maintaining the calcification conspicuity and architectural distortion visibility required for accurate BI-RADS assessment.

Microcalcifications—a key indicator of ductal carcinoma in situ (DCIS)—present a unique challenge in DBT due to their small size and the limited angular sampling. To enhance their visibility, radiologists rely on slab reconstruction techniques such as MIP, which collapses 5-10 consecutive slices into a single image. This aggregates the signal from a calcification cluster distributed across multiple slices, making the morphology and distribution pattern (e.g., linear branching vs. grouped punctate) more conspicuous for malignancy risk stratification.

The iterative reconstruction algorithms used in DBT introduce unique artifacts that AI detection models must be trained to ignore. These include out-of-plane blurring from high-contrast objects (e.g., surgical clips), limited-angle aliasing that creates ripple patterns, and boundary ringing at tissue-air interfaces. A robust CADe/CADx system must differentiate these reconstruction-induced signals from true radiological findings such as spiculated masses or architectural distortion to maintain high specificity and avoid false positive marks.

A standard DBT study generates 200-300 individual slices per view, significantly increasing radiologist interpretation time compared to a single 2D image. Studies indicate that reading a DBT exam takes approximately 2x longer than FFDM. This workflow burden has driven the adoption of AI-driven concurrent reading aids and worklist prioritization algorithms. These tools pre-screen the volumetric data, highlight slices containing suspicious Regions of Interest (ROIs), and reorder the reading queue so high-suspicion cases are interpreted first, mitigating reader fatigue.

An AI system designed to automatically mark suspicious regions in a medical image to reduce observational oversights. In DBT, CADe algorithms operate on the reconstructed slice data or Maximum Intensity Projection (MIP) images to identify masses and calcifications. Modern deep learning CADe significantly outperforms traditional feature-based systems by learning hierarchical representations directly from volumetric data.

A 2D visualization technique that projects the brightest voxels from a slab of DBT slices into a single image. MIPs serve as a critical bridge between 3D acquisition and 2D review, allowing radiologists to rapidly localize calcifications and suspicious masses without scrolling through every slice. AI models often use MIPs as an input channel to reduce computational complexity while preserving diagnostic information.

A post-processing AI technique designed to suppress erroneous marks generated by a detection model. In DBT workflows, false positive reduction is essential because the increased anatomical detail of tomosynthesis can cause algorithms to flag normal superimposed tissue or blood vessels as lesions. This step directly improves specificity and reduces unnecessary recall rates.

A statistical analysis curve that plots the true positive detection rate against the average number of false positives per image. FROC is the standard metric for evaluating localization performance in DBT CADe systems. Unlike ROC analysis, FROC accounts for the spatial component of detection, penalizing algorithms that correctly identify a lesion but mark multiple false locations elsewhere in the volume.

An algorithmic process that geometrically links findings across the Craniocaudal (CC) and Mediolateral Oblique (MLO) views. In DBT, multi-view correlation leverages the 3D coordinate system to triangulate a lesion's position, confirming true abnormalities and suppressing artifacts visible in only one projection. This technique is a powerful lever for reducing false positives in screening workflows.

A clinical integration model where AI displays detection marks simultaneously while the radiologist interprets the DBT slices in real-time. This contrasts with a second-reader workflow where AI results are shown only after the radiologist's initial interpretation. Concurrent reading reduces interpretation time by directing attention to suspicious regions immediately, but requires careful UI design to avoid automation bias.