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

PriorNet: Temporal CAD refinement for digital breast tomosynthesis

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11427939

1.0

Yinhao Ren, Duke University, 2424 Erwin Rd, Studio #302, Durham, NC 27705; email: ude.ekud@ner.oahniy

Supported/funded by iCAD Inc. under sponsored research collaboration agreement (SPS#264557).

Retrospective study approved by sponsor’s IRB; HIPAA compliant; informed consent waived.

Cascaded deep learning module atop prior IMR-Net. PriorNet includes (1) coarse center-of-mass alignment of current and 1-year-prior DBT volumes, (2) temporal matching network (twin/siamese CNN) to pair current and prior detections, and (3) growth estimation network (twin/siamese CNN) to estimate lesion growth. Outputs pooled to case-level overall score as weighted sum of detection and growth across four views.

Not publicly released; developed under sponsored research with iCAD Inc.

Improves breast cancer lesion detection performance in DBT screening by incorporating temporal change, with higher AUC and improved specificity at high sensitivity.

Clinical decision support systems; workflow optimization during screening DBT interpretation using prior comparisons.

Assists radiologists by automatically detecting lesion candidates, matching with prior exam, estimating growth, and producing case-level scores; does not replace clinical judgment.

Computer-aided detection for screening digital breast tomosynthesis in adult patients with a 1-year prior DBT available; intended for use in breast imaging settings with Hologic DBT systems.

Current and 1-year-prior bilateral DBT examinations (CC and MLO views) from Hologic systems; detection candidate patches of 700×700×7 pixels used after coarse alignment.

Use with both current and 1-year prior DBT exams. System computes lesion-level suspiciousness and growth, then aggregates to a case-level overall score. High-suspicion lesions may be considered even if temporal growth is low per study design.

Dataset included only 1-year priors (clinical reference for negatives often requires 2-year stability); imbalance in breast density categories and cancer subtype profiles; case-level pooling may not mirror lesion-level clinical recommendations; current-prior matching may occasionally mismatch nearby detections; all data from Hologic systems; external validation limited to one external site.

Incorporate temporal information from prior DBT alongside ipsilateral matching to enhance CAD detection performance and reduce false positives, especially at high-sensitivity operating points.

Implemented in TensorFlow 2.5 (Python 3.7); trained with Adam optimizer (lr=1e-4), standard augmentations (brightness, scaling, cropping, rotation); developed and tested on NVIDIA RTX 3090 Ti GPU; training/validation/testing split by site: 2604/2599/2861 cases.

Per examination computation: baseline IMR-Net ~21.7 seconds; PriorNet adds <4 seconds overhead.