The FDA cleared an AI tool that flags breast-cancer lesions on ultrasound. The accuracy numbers behind it aren't published yet.

The FDA cleared a new AI tool for reading breast ultrasounds on July 28: 510(k) K260303, granted to See-Mode Technologies for the See-Mode Augmented Reporting Tool, Breast — sold in the US as DeepHealth Breast Ultrasound. DeepHealth is a subsidiary of RadNet, the outpatient imaging chain, which acquired See-Mode's Australian parent company as part of its push into AI-assisted radiology. The software automates detection and characterization of suspicious lesions on breast ultrasound images — shape, orientation, margin, echo pattern, and posterior features, described using terminology aligned with the American College of Radiology's BI-RADS system — and drafts a structured report. The radiologist reviewing the scan retains final sign-off; the software does not issue a diagnosis on its own.

A 510(k) clearance is a narrower regulatory bar than it can sound like. It means the FDA agrees the device is *substantially equivalent* to an already-cleared device of the same type — in this case, See-Mode's own prior clearances — not that the agency independently verified the specific accuracy numbers RadNet is now publicizing. Those numbers come from RadNet's own multi-reader, multi-case study: 16 US board-certified radiologists reading the same case set with and without the software, which RadNet reports produced greater than 98% lesion-localization accuracy, an 8-percentage-point improvement in cancer-detection sensitivity, and a 37% reduction in radiologist interpretation time. As of this writing, that study has not appeared in a peer-reviewed journal — the figures are RadNet's own characterization of its own trial, not yet independently checked by outside reviewers.

Breast ultrasound itself is already a common supplemental exam — used alongside mammography for patients with dense breast tissue, where mammograms alone are known to miss more findings, and for following up on something a mammogram flagged. Reading ultrasound images has historically been more subjective than reading a mammogram, since the sonographer's positioning and the radiologist's real-time interpretation both shape what gets reported, which is precisely the kind of variability an automated first-pass characterization tool is aimed at reducing. That context is why a clearance like this one matters beyond its own numbers: it's not introducing AI into a new corner of medicine so much as targeting the specific step in an already-common exam where human interpretation varies most.

The billing code RadNet is counting on is also worth reading carefully. Category III CPT codes, the type covering quantitative ultrasound tissue characterization here, are temporary tracking codes for emerging technology — they let a service be billed and studied, but they don't carry the same payment guarantee as an established Category I code, and individual insurers decide separately whether and how much to pay against them. RadNet's 700,000-study estimate is a ceiling on studies that could qualify, not a confirmed reimbursement figure from any specific payer.

Not RadNet's first anatomy

SMART-B is the third device in a pattern, not a one-off. See-Mode's FDA record shows the same augmented-reporting approach applied to a new body part roughly every two years: a vascular-ultrasound tool in 2020, a thyroid-ultrasound tool in 2024, and now breast ultrasound in 2026. Separately, DeepHealth's own mammography-AI line — Saige-Dx and Saige-Density, acquired from an earlier startup — has been through several of its own clearances and updates over the same stretch. RadNet is assembling a portfolio that covers more of the imaging a single outpatient network runs, one modality at a time, rather than betting on a single flagship product.

The rollout plan is where the clearance turns into a business decision rather than just a regulatory milestone. RadNet intends to deploy DeepHealth Breast Ultrasound across its network of more than 400 outpatient imaging centers by the end of 2026, and estimates more than 700,000 annual breast-ultrasound studies could qualify for reimbursement under an existing Category III CPT code covering quantitative ultrasound tissue characterization. That reimbursement pathway matters as much as the clearance itself: a cleared device with no billing code attached to its output is a much harder sell to a radiology practice than one that fits an existing payment mechanism.

SMART-B arrives alongside a broader wave of FDA activity in AI-assisted imaging this year — Aidoc and Cognita have each received FDA breakthrough-device designations for tools that read chest X-rays and draft reports, a status that fast-tracks FDA review but is not itself a clearance. The pattern across all of these tools is consistent: automate the parts of a radiologist's report that are structured and repetitive, leave the actual call to a licensed physician, and let the reimbursement code do the work of making adoption an economic decision rather than a purely clinical one. Whether the specific accuracy gains RadNet is citing hold up once independent researchers can review the underlying study is the open question the clearance itself doesn't answer.