mg/dbt

Breast Tomosynthesis

DBT read: lesions located in 3D with their best slice, exam score, recall and BI-RADS.

Coming soonMGMammography & Tomosynthesis

This model is not available to run yet. Vote to show interest and get an email when it opens.

Overview

What it does

Breast Tomosynthesis reads a digital breast tomosynthesis (DBT) study: one reconstructed stack of thin slices for each view, usually with synthesized 2D images and sometimes with conventional FFDM views. DBT finds more cancers and causes fewer recalls than 2D mammography, but a radiologist has to scroll through 50 to 90 slices per view, which roughly doubles the reading time.

The run searches every slice of every stack. Each mass, area of architectural distortion and calcification cluster is returned as a 3D box with its slice range, the single best slice to look at, and its own score. The exam-level read adds a suspicion score, a recall recommendation and a BI-RADS assessment, so the radiologist can open the study at the slices that matter.

Intended use

Decision support for the radiologist reading DBT in screening or diagnostic work: a second read that points to the slices and regions that matter, and a structured exam assessment with a recall recommendation. Every lesion comes with its location in the stack.

Who it is for

Breast and general radiologists reading tomosynthesis, screening programmes that have moved from 2D to DBT, and breast clinics. Developers can call the run from a PACS or screening worklist to pre-mark key slices before the reader opens the study.

Inputs and protocol

Accepted input

DICOM Breast Tomosynthesis Image objects (one stack per view) and Breast Projection X-Ray images.
Vendor-private tomosynthesis objects, which are decoded into standard stacks before reading.
Synthesized 2D and FFDM views of the same examination, uploaded in the same study.
Priors in prior_files, as DBT or 2D examinations.

A DBT study is large, typically 1 to 3 GB, with 50 to 90 slices per view at about 1 mm spacing.

Requirements

Requirement	Why
A tomosynthesis stack for each standard view	Lesions are detected in 3D and matched between CC and MLO
`ImageLaterality` and view codes	Each stack is assigned to the correct breast and projection
Slice spacing and pixel spacing	Slice ranges and lesion sizes in mm
`BreastImplantPresent`	Implant and implant-displaced views are recognised

Image intensities are normalised per vendor before reading; incomplete studies are flagged.

Optional context

Screening or diagnostic exam_context, age, risk factors and the clinical indication in clinical_context.

Outputs and standards

The result

A JSON result with the exam suspicion score, recall decision, BI-RADS assessment and a lesion list carrying slice coordinates.

Section	Content
Lesions in 3D	Masses, architectural distortion and calcification clusters as 3D boxes with slice range, best slice, laterality, clock-face location and per-lesion score
Draft report	Opt-in (`options.report`): an English narrative that interprets the result for the reader (findings, impression, limitations); every number is checked against the findings

Standards

DICOM: Mammography CAD SR (TID 4000) with slice coordinates for each finding; GSPS overlays on the key slices; Key Object Selection marking the key DBT slices; display through the IHE AI Results (AIR) profile.
Reporting: ACR BI-RADS v2025 Manual for assessment categories and the lexicon used in lesion descriptors.
Terminology: RadLex BI-RADS terms and SNOMED CT.

Relation to the 2D model

Breast Tomosynthesis shares its lesion lexicon, location fields and assessment logic with Mammogram (2D), so DBT and 2D reads of the same patient line up when the reader compares them over time.