Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/model.json

Name

Transformer-based Deep Neural Network for Breast Cancer Classification on Digital Breast Tomosynthesis Images

Link

https://dx.doi.org/10.1148/ryai.220159

Indexing

Keywords: breast cancer, digital breast tomosynthesis, DBT, transformer, TimeSformer, convolutional neural network, heat map, classification

Content: BR

RadLex: RID10359, RID39055, RID34265, RID29922, RID34246, RID43364, RID34261

SNOMED: 254837009, 309587003

Author(s)

Weonsuk Lee

Hyeonsoo Lee

Hyunjae Lee

Eun Kyung Park

Hyeonseob Nam

Thijs Kooi

Organization(s)

Lunit

Version

1.0

Contact

oi.tinul@2675koesnowi

Funding

Supported by funds secured by Lunit.

Ethical review

Study data were retrospectively collected in compliance with HIPAA; all examinations were de-identified via the Safe Harbor method, and institutional review board approval was not required.

Date

Published: 2023-05-10

References

[1] Lee W, Lee H, Lee H, Park EK, Nam H, Kooi T. "Transformer-based Deep Neural Network for Breast Cancer Classification on Digital Breast Tomosynthesis Images". Radiology: Artificial Intelligence. 2023;5(3):e220159. 2023-05-10. doi:10.1148/ryai.220159. PMID: 37293346. PMCID: PMC10245183.

Model

Architecture

Transformer-based model using TimeSformer with divided space-time attention operating on neighboring DBT sections, with a 2D CNN backbone for per-section feature extraction and an aggregation network for section-level score and pixel-level heat map.

Clinical benefit

Improved breast cancer detection on DBT with reduced false positives and fewer missed cancers compared with a per-section baseline; potential to aid radiologists and reduce workload.

Clinical workflow phase

Clinical decision support systems; potential triage/detection aid during image interpretation.

Decision threshold

Evaluation reported at operating points: sensitivity at fixed specificity 0.8 and specificity at fixed sensitivity 0.8.

Degree of automation

Fully automated image analysis producing section-level malignancy likelihoods and heat maps to support, not replace, clinician decision-making.

Indications for use

Research-stage system for detecting malignant lesions in adult breast digital breast tomosynthesis (four-view) studies acquired on Hologic scanners in radiology settings.

Input

Four-view digital breast tomosynthesis stacks; subsets of neighboring sections are sampled for context during training; full stack evaluated at inference.

Instructions

Model processes DBT sections with a 2D backbone, applies transformer-based interaction across neighboring sections, and aggregates features to produce section-level malignancy scores and pixel-level heat maps; during testing, predictions are made for the entire DBT stack.

Limitations

Evaluated on a single-institution test set; all data from a single manufacturer (Hologic); U.S.-only population; generalization to other vendors/sites/populations not yet demonstrated.

Output

CDEs: RDE65, RDE1586

Description: For each DBT section: malignancy likelihood score and a pixel-level heat map indicating likelihood of malignant lesion.

Recommendation

For research use; additional external validation across institutions, vendors, and populations is needed before clinical deployment.

Reproducibility

Not reported.

Sustainability

TimeSformer requires ~25% of the FLOPS of a comparable Conv3D model; latency comparable to single-section baseline on reported hardware/software (exact times hardware-dependent).

Use

Intended: Detection and diagnosis

Out-of-scope: Other

User

Intended: Radiologist, Subspecialist diagnostic radiologist