A Deep Learning Model for Breast Shear Wave Elastography to Improve Breast Cancer Diagnosis (INSPiRED 006): An International, Multicenter Analysis

Background: Breast shear wave elastography (SWE) has been evaluated to enhance B-mode breast ultrasound performance. Although multi-center trials have shown potential benefits for patients with BI-RADS 4(a) breast masses, its adoption is limited due to the lack of a prospectively validated velocity cutoff for malignancy. This study aims to develop and validate a deep learning (DL) model using SWE images for BI-RADS 3 and 4 breast masses and compare its performance to human experts using B-mode ultrasound.

Methods: Data from an international, multicenter trial (NCT02638935) were used to assess SWE's diagnostic performance in women with BI-RADS 3/4 breast masses across 12 institutions in 7 countries. Images from 11 study sites were used to train a DL model with an EfficientNetB1 backbone, and images from the 12^th study site served as an external validation set. Performance metrics included sensitivity and false-positive rates and AUROC.

Results: The study included 924 patients (4026 images) for model development and 194 patients (562 images) for external validation. Breast cancer was diagnosed in 29.2% of the development set and 24.2% of the validation set. The SWE DL model achieved an AUROC of 0.94 in the external validation set. Compared to B-mode ultrasound, the SWE DL model significantly reduced the false-positive rate (20.41% vs. 53.81%, 62.07% reduction) with comparable sensitivity (97.9% vs. 98.1%).

Conclusion: The SWE-DL model showed equivalent accuracy to human experts in detecting malignancies while significantly reducing unnecessary biopsies. Future studies should evaluate its integration into multi-modal breast cancer diagnostics in a prospective setting.

Publikationsverlauf

Artikel online veröffentlicht:
04. Juni 2025