AI System That Combines Ultrasound, Mammography and MRI Could Slash Unnecessary Breast Biopsies

A new artificial intelligence system that reads multiple types of breast imaging together could help substantially reduce the number of women who undergo invasive needle biopsies for benign lesions, while maintaining high cancer detection rates.

The system, called BINDS, was developed by a team led by researchers at ShanghaiTech University and described this week in Nature Biomedical Engineering. It integrates ultrasound, mammography and MRI, and is designed to mirror the two-step workflow that radiologists actually use: an initial assessment with widely available imaging, followed by a more detailed look with MRI when the first results are uncertain.

The researchers built and validated BINDS using one of the largest multimodal breast imaging datasets assembled to date — 27,048 participants from eight Chinese hospitals and seven public datasets. In a reader study involving radiologists from Yunnan Cancer Hospital, BINDS delivered several key results.

When working with just one type of imaging, the AI’s diagnostic accuracy was on par with that of experienced radiologists. However, its advantage emerged clearly when multiple modalities were combined. Using a two-stage approach that starts with ultrasound and mammography, then selectively adds MRI for uncertain cases, BINDS achieved an area under the receiver operating characteristic curve (AUROC) of 0.973 on an internal test set. In a reader study using a subset of 208 cases from that cohort, the AI's performance surpassed both junior and senior radiologists.

Crucially, when radiologists were given the AI’s predictions and attention heat maps that highlighted suspicious image regions, their ability to avoid unnecessary biopsies improved markedly. For senior radiologists performing a three-modality assessment, the number of benign biopsies dropped by 32.4%, while the biopsy rate for malignant lesions remained unchanged.

Breast cancer diagnostic pathways generate many false positives, particularly among lesions initially classified as BI-RADS 4 — a category that carries a malignancy risk anywhere from 2% to 95%. Current guidelines recommend biopsies for these patients, but many ultimately prove benign. Reducing those biopsies without missing cancers would spare women physical discomfort, anxiety and healthcare costs.

BINDS addresses a practical reality: not all patients receive all three imaging modalities. Its flexible design can handle any combination of available inputs — ultrasound alone, mammography plus ultrasound, or all three — making it potentially usable across clinics with different resources. The system also indicates which modality is driving its decision in any given case, an interpretability feature meant to build radiologist trust.

This is a retrospective study. The in-house training and test data came entirely from Chinese medical centres, which may limit how well the model generalizes to other populations and imaging equipment. The radiology–pathology alignment component, which helps the system learn tissue-relevant features, used pathology images from only a single hospital. The authors note that a prospective, multi-centre study is needed to confirm whether the reduction in benign biopsies holds up in real-world clinical use and to show a definitive patient benefit.

If validated prospectively, such systems could serve as a safety net — flagging cases where a biopsy might be safely deferred, or, conversely, prompting a closer look when early imaging findings are ambiguous. The team also suggests the framework could be extended to incorporate advanced ultrasound techniques, potentially reducing reliance on more expensive MRI even further.

---

Reference
Li, Y. et al. A deep learning system for non-invasive breast cancer diagnosis with multimodal data. Nat. Biomed. Eng. (2026). https://doi.org/10.1038/s41551-026-01388-9