DeepHealth breast ultrasound FDA clearance: AI cuts reads 37%
DeepHealth breast ultrasound FDA clearance came on July 30, 2026, when the RadNet subsidiary announced 510(k) clearance for an AI tool that detects, describes and reports breast lesions during scanning and reading. In a study with 16 radiologists, the company says the tool cut interpretation time by 37%. For imaging providers, it targets an operator-dependent exam that already has a reimbursement code, across an estimated more than 700,000 RadNet studies a year.
Why breast ultrasound is hard to standardize
Breast ultrasound depends heavily on the person holding the probe and the person reading the images. Approximately 40% of women undergo the exam at some point in their lives, according to Dr. Jason McKellop, Medical Director of Women's Imaging for RadNet California.
"It is a highly complex, operator-dependent examination, which can lead to significant variability in image acquisition, interpretation and reporting." — Dr. Jason McKellop, RadNet California
For an imaging business, variability means slower exams, uneven reports and more time spent by sonographers on manual documentation.
What the FDA cleared for DeepHealth
The FDA granted 510(k) clearance to DeepHealth Breast Ultrasound, an AI-powered tool that automates lesion detection, characterization and reporting. DeepHealth, based in Somerville, Massachusetts, is a wholly owned subsidiary of RadNet, which trades on Nasdaq as RDNT.
With clearance, the product is commercially available to customers in the United States. DeepHealth says it now forms part of the industry's most comprehensive breast platform, alongside its AI tools for mammography, density assessment and breast cancer risk prediction.
How the AI works during a scan and a read
The tool covers four steps of the exam. During detection, it helps the interpreting physician find suspicious soft-tissue lesions in standard ultrasound images. During characterization, it describes each lesion's shape, orientation, margin, echo pattern and posterior features in line with ACR BI-RADS.
It then drafts a radiology report of the key findings and impressions, while the radiologist keeps control of the final assessment. For sonographers, it extracts and organizes lesion measurements and other findings in a standard format, with the aim of reducing manual documentation and overall exam time.
What the study results show
DeepHealth reports three results from a multi-reader multi-case study with 16 U.S. board-certified radiologists at imaging centers and hospitals. The company says the tool showed greater than 98% accuracy in localizing breast lesions, improved sensitivity for breast cancer detection by 8% and reduced radiologist interpretation time by 37%.
RadNet says it has also validated the tool in live clinical settings under regulated research protocols. These figures are the company's own; the press release does not publish the full study.
What it changes for imaging centers
The business case rests on payment and throughput. Customers can pursue reimbursement under an existing Category III CPT code for quantitative ultrasound tissue characterization. RadNet plans to implement the tool across its network by the end of this year, with an estimated more than 700,000 breast ultrasound studies annually that may be eligible for reimbursement.
RadNet chief executive Dr. Howard Berger said the company anticipates both revenue and cost savings contributions from the roll-out in the second half of the year. That is a forecast. The Digital Health segment that sells DeepHealth products reported annual recurring revenue of $105.5 million at June 30, 2026, up 97.2% from a year earlier.
How DeepHealth sells beyond RadNet
DeepHealth's products are already sold outside its parent. RadNet says external customers now generate approximately 63% of Digital Health revenue, and that new business signed in the second quarter had a total contract value of approximately $21 million, mostly with hospitals and health systems. RadNet runs 442 owned and operated outpatient imaging centers, which gives the ultrasound tool a large first home before it reaches outside buyers.
Questions and answers
What does DeepHealth Breast Ultrasound do?
It assists sonographers and radiologists by detecting suspicious lesions, characterizing them in line with ACR BI-RADS, drafting a radiology report and organizing lesion measurements. The radiologist keeps control of the final assessment.
Can providers get reimbursed for AI breast ultrasound?
Yes, according to DeepHealth. Customers in the United States can pursue reimbursement under an existing Category III CPT code for quantitative ultrasound tissue characterization.
When will RadNet use the tool across its centers?
DeepHealth says the solution will be implemented across RadNet's network of centers by the end of this year.
MOZGII view
How we would build it
Robots and AI help most in medicine where work is repeated, measured and easy to check: positioning, measuring, documenting and flagging what a specialist should look at first. Breast ultrasound fits that description. The images are standard, the descriptions follow ACR BI-RADS, and the radiologist already signs every report.
If we were building a tool like this, we would start with the data: ultrasound images, the sonographer's measurements, the radiologist's final BI-RADS description and, where it exists, the biopsy result that confirms or rules out cancer. The model would draft the lesion description and the report, and the radiologist would approve or correct every draft. Each correction would go back into a review set, so the team can see where the model and the specialists disagree. We would keep the final assessment with the physician, as DeepHealth does.
It would break where the images differ from the training data: a new ultrasound machine, a new site, or a sonographer with a different scanning habit. Results can drift quietly after a roll-out across many centers. The one metric we would watch is how often radiologists change the AI's draft report, tracked by site and by machine, because a rising edit rate is the earliest sign that the model no longer fits the clinic.
Next, we would build a scan-quality check that tells the sonographer, during the exam, when an image is not good enough to read.
Sources
Where the facts in this story come from.