Research
Turning general AI into radiology specialists.
What we are testing
Manuscript in preparationSelective fine-tuning
Retrain only the parts of a model that matter for radiology, rather than the whole network, so a new model can be adapted quickly and at low cost.
Specialist layers on top
Add lightweight layers on top of a general vision-language model to strengthen domain-specific perception, leaving the base model intact.
Keeping pace with new models
A repeatable recipe that carries radiology skill over to each new open model as it is released.
Why it matters
Percent growth, United States
Source: AuntMinnie, September 2026.
Annual attrition, United States
Source: ACR Bulletin, February 2026.
Programs
- Manuscript
Re-adapting new models fast
Making each new open model radiology-ready, quickly and cheaply.
- Active
Grounded reporting
Every finding cites the image it came from.
- Active
Learning from radiologist edits
One correction improves every future read.
- Active
Detector-assisted reading
A small detector beside the writer, for subtle findings.
- Active
Ultrasound worksheet fusion
Worksheet, sketch and calipers read alongside the images.
- In research
X-ray and head CT readers
The same grounded drafting, beyond ultrasound.
From the field
RSNA 2026 technical exhibits spotlight medical imaging innovation
Radiologist pipeline trails increase in training positions
How I Do It: understanding and leveraging generative AI for chest radiology reporting
FDA clears AI-enabled ultrasound for high-volume clinical settings
Aidoc's First Read X-ray reporting software gets FDA Breakthrough Device designation
Radiology gets 68 new FDA-cleared algorithms
Papers we build on
- Efficiency and quality of generative AI-assisted radiograph reportingHuman-in-the-loopJAMA Network Open · 2025
- AI assisting junior radiologists in thyroid ultrasound: a multicenter prospective studyThyroid USBMC Medicine · 2024
- From ACR O-RADS 2022 to explainable deep learning in ovarian massesPelvic USarXiv preprint · 2025
- Multimodal ultrasound radiomics and deep learning for O-RADS 4-5 adnexal massesPelvic USCancer Imaging · 2025
- MedGemma technical reportReport generationarXiv preprint · 2025
- EditGRPO: reinforcement learning with post-rollout edits for chest X-ray report generationHuman-in-the-loopAACL 2025 · 2025