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

NYU Breast Cancer Screening Classifier (NYU1/NYU2) retrained on BreastScreen SA data

https://github.com/jamesjjcondon

1.0

ua.ude.edialeda@remlap.elyl

Authors declared no funding for this work.

Approved by local institutional review board (HREC/16/229, R20160601), with a waiver of consent for the retrospective use of de-identified clinical data.

Convolutional Neural Network. Two variants evaluated: NYU1 (image-only, four-view input) and NYU2 (image + patch-level benign/malignant heatmaps generated by a separate CNN). Network architecture unchanged from Wu et al.

Study code: https://github.com/jamesjjcondon (repository for transparency; relies on private data and not currently operational). Original NYU model/code: https://github.com/nyukat/breast_cancer_classifier

Intended to assist in classifying screening mammograms for malignancy to potentially support radiologist performance; in this study, evaluated for generalizability and impact of transfer learning.

Clinical decision support systems (research evaluation; not deployed clinically).

For the best-performing retrained NYU2 model, operating point selected to achieve validation subset sensitivity above 90%.

Decision support; not suitable as a stand-alone reader in the studied local population due to low specificity.

Binary classification of malignancy (invasive cancer or DCIS) versus no malignancy/benign on screening full-field digital mammography in adult women within a population screening program setting.

Four-view screening mammograms (bilateral CC and MLO). For NYU2, additional inputs are benign and malignant patch-level heatmaps generated by a CNN trained on hand-drawn segmentations from the NYU dataset.

Preprocess to match NYU pipeline; fixed input size required; right-sided images horizontally flipped for training/validation/testing; per-individual malignancy score taken as the maximum image-wise score across views; for transfer learning, initialize with NYU weights and retrain on local training subset.

Out-of-the-box performance decreased on local Australian data; improved with transfer learning contingent on access to original model weights (which are not generally available). Heatmap-generating CNN was not retrained due to lack of local segmentations. Differences in vendor devices, image size, and image characteristics may affect generalizability. Interval cancer data not available; breast density not routinely collected. In the studied cohort, specificity remained below typical clinical standards.

Models should be tested on and adapted with local data prior to clinical use; provision of complete model details and weights facilitates adaptation.

All code used for this study is available at https://github.com/jamesjjcondon; repository depends on private BSSA data and is not currently operational for replication.