
Stanford - Radiology AI | RadAI @ HOPPR | Previous: ML @HuggingFace, Academic staff - Research @Stanford University, @StanfordAIMI Affiliate
Palo Alto
Joined July 2017
- Tweets306
- Following939
- Followers1.2K
- Likes219
#RadiologyAI Our report drafting solution is deployed in 300 sites over 3 countries.
Here are ๐๐ต๐ฒ ๐ผ๐ฝ๐ฒ๐ป ๐พ๐๐ฒ๐๐๐ถ๐ผ๐ป๐ we faced when building VLMs that scale across healthcare systems, radiologists of different nationalities, reporting styles, clinical use cases, and varying finding prevalence.
This effort created ๐๐ป๐ฝ๐ฟ๐ฒ๐ฐ๐ฒ๐ฑ๐ฒ๐ป๐๐ฒ๐ฑ ๐ ๐ ๐ฐ๐ต๐ฎ๐น๐น๐ฒ๐ป๐ด๐ฒ๐: how can a single foundation model satisfy radiologists when one reports aortic knob prominence 90% of the time, while their colleague next door reports it 0% of the time?
After all, we have all adapted to using the same ChatGPT or Claude. Radiologists, however, are not ready to adapt to a single, standardized threshold for reporting findings. ๐ง๐ต๐ฒ ๐บ๐ผ๐ฑ๐ฒ๐น ๐บ๐๐๐ ๐ฎ๐ฑ๐ฎ๐ฝ๐ ๐๐ผ ๐๐ต๐ฒ๐บ, ๐ป๐ผ๐ ๐๐ต๐ฒ ๐ผ๐๐ต๐ฒ๐ฟ ๐๐ฎ๐ ๐ฎ๐ฟ๐ผ๐๐ป๐ฑ.
๐๐ป๐ฑ ๐๐ต๐ฎ๐ ๐ถ๐ ๐ฝ๐ฒ๐ฟ๐ณ๐ฒ๐ฐ๐๐น๐ ๐ณ๐ถ๐ป๐ฒ. But to do so, you need to understand many overlooked element in the NLP/NLG scientific literature:
- How do you control a model's conservatism or verbosity through metadata?
- How do you define an operating point, a decision boundary in language space, that determines whether a finding should be reported, especially when generating an entire radiology report section?
- How do you train a decoder to predict whether a radiologist will sign off on a generated report?
- And most importantly, how do you truly understand your end user: the radiologist?
At Eden, our technology stack is already deployed across ๐บ๐ผ๐ฟ๐ฒ ๐๐ต๐ฎ๐ป ๐ฎ,๐ด๐ฌ๐ฌ ๐ถ๐บ๐ฎ๐ด๐ถ๐ป๐ด ๐ฑ๐ฒ๐ฝ๐ฎ๐ฟ๐๐บ๐ฒ๐ป๐๐ ๐ถ๐ป ๐ฎ๐ฌ ๐ฐ๐ผ๐๐ป๐๐ฟ๐ถ๐ฒ๐, giving us a unique opportunity to solve these challenges at scale.
#RadiologyAI New proceedings alert ๐จ . Finally, someone does the painstaking groundwork. ๐ก๐๐โ๐ ๐๐ง-๐๐ฒ๐ป๐ฐ๐ต ๐ฝ๐๐๐ ๐๐ผ๐ฟ๐ฑ๐ ๐ผ๐ป ๐ฎ๐ฌ,๐ฏ๐ฏ๐ฑ ๐๐ง ๐น๐ฒ๐๐ถ๐ผ๐ป๐ (๐ ๐๐๐๐ ๐ฎ๐ฌ๐ฎ๐ฒ). ๐ข๐ป ๐ถ๐๐ ๐ณ๐ผ๐๐ฟ-๐ฐ๐ต๐ผ๐ถ๐ฐ๐ฒ ๐ฒ๐
๐ฎ๐บ, ๐๐ฃ๐ง-๐ฐ๐ฉ ๐๐ฐ๐ผ๐ฟ๐ฒ๐ ๐ฎ๐ด%.
CT-Bench is an exam for AI on single CT lesions: which of these four descriptions fits this lesion? Which image matches this description? How big is it?
Picking at random scores 25%. GPT-4V scores ๐ฎ๐ด%. It is a 2023โ2024 model; nothing newer was tested.
๐๐ป๐ฑ ๐ฟ๐ฎ๐ฑ๐ถ๐ผ๐น๐ผ๐ด๐ถ๐๐๐? ๐ง๐๐ผ ๐๐ฒ๐ป๐ถ๐ผ๐ฟ ๐ฟ๐ฎ๐ฑ๐ถ๐ผ๐น๐ผ๐ด๐ถ๐๐๐ ๐ด๐ผ๐ ๐ฎ๐ฏ๐ผ๐๐ ๐ด ๐ถ๐ป ๐ญ๐ฌ ๐ฟ๐ถ๐ด๐ต๐, ๐ผ๐ป ๐ญ๐ฌ๐ฌ ๐ฐ๐ฎ๐๐ฒ๐.
โ The best AI in the paper, a model trained on the released lesions, gets ๐ฑ๐ฌ% overall.
โ Draw a box around the lesion and that model goes from 37% to 62%. The senior radiologists go from about 7 in 10 to about 9 in 10. Knowing where to look matters for both.
๐ช๐ต๐ฎ๐ ๐๐ผ๐ ๐ฐ๐ฎ๐ป ๐๐๐ฒ ๐๐ผ๐ฑ๐ฎ๐:
โ ๐ฎ๐ฌ,๐ฏ๐ฏ๐ฑ ๐๐ง ๐น๐ฒ๐๐ถ๐ผ๐ป๐, each with a one-sentence description, a size and a box, plus 2,850 exam questions with the paperโs scores to compare against. On Kaggle (link in the first comment).
โ The licence is CC BY-NC-SA 4.0: free for research, ๐ป๐ผ ๐ฐ๐ผ๐บ๐บ๐ฒ๐ฟ๐ฐ๐ถ๐ฎ๐น ๐๐๐ฒ.
The lesson: before trusting a model with a whole report, test it lesion by lesion: ๐๐ต๐ฒ๐ฟ๐ฒ ๐ถ๐ ๐ถ๐, ๐๐ต๐ฎ๐ ๐ถ๐ ๐ถ๐, ๐ต๐ผ๐ ๐ฏ๐ถ๐ด ๐ถ๐ ๐ถ๐.
Public CT data that ties a sentence to a single lesion is rare, and this set is a download away.
Full breakdown in the first comment ๐
Blog post: jbdel.github.io/research-bloโฆ
Paper: melba-journal.org/papers/202โฆ
Data: kaggle.com/datasets/cd1661d6โฆ
#RadiologyAI My highlight of #MICCAI2026 ๐๐๐ป๐ถ๐โ๐ ๐ฆ๐ฃ๐๐-๐๐ซ๐ฅ ๐ด๐ฟ๐ฎ๐ฑ๐ฒ๐ ๐๐ ๐ซ-๐ฟ๐ฎ๐ ๐ฟ๐ฒ๐ฝ๐ผ๐ฟ๐๐ ๐ณ๐ถ๐ป๐ฑ๐ถ๐ป๐ด ๐ฏ๐ ๐ณ๐ถ๐ป๐ฑ๐ถ๐ป๐ด. ๐ง๐ต๐ฒ ๐ฐ๐ฎ๐๐ฐ๐ต: ๐ผ๐ป ๐บ๐ผ๐๐ ๐ธ๐ถ๐ป๐ฑ๐ ๐ผ๐ณ ๐ณ๐ถ๐ป๐ฑ๐ถ๐ป๐ด, ๐๐ต๐ฒ ๐ฏ๐ฒ๐๐ ๐๐ ๐ถ๐ ๐บ๐ผ๐๐๐น๐ ๐๐ฟ๐ผ๐ป๐ด.
SPEC-CXR asks four questions about every finding: is it there, in the right place, with the right severity, and compared correctly with the last X-ray?
Tubes and devices get found. But on ๐ต ๐ผ๐ณ ๐ญ๐ฑ kinds of finding, the best model, Med-Gemini, either misses at least 4 in 5 real cases or gets at least 4 in 5 of its calls wrong.
๐ช๐ต๐ฎ๐ ๐๐ผ๐ ๐ฐ๐ฎ๐ป ๐๐๐ฒ ๐๐ผ๐ฑ๐ฎ๐:
โ The code (link in the first comment), free for non-commercial research.
โ The checklist of 122 findings and diagnoses, as plain files you can reuse (same licence) for your own per-finding table.
โ One rule: name your AI judge and keep it fixed. Across the paperโs eight judges, agreement with radiologists ranges from ๐ฌ.๐ฒ๐ฌ ๐๐ผ ๐ฌ.๐ณ๐ด.
Caveat: the finding-level grades were checked on only 50 report pairs, and the released code runs a different version of the pipeline from the one the paper validated.
The lesson: ๐ผ๐ป๐ฒ ๐๐ฐ๐ผ๐ฟ๐ฒ ๐ฝ๐ฒ๐ฟ ๐ฟ๐ฒ๐ฝ๐ผ๐ฟ๐ ๐ฐ๐ฎ๐ปโ๐ ๐๐ฎ๐ ๐๐ต๐ถ๐ฐ๐ต ๐ณ๐ถ๐ป๐ฑ๐ถ๐ป๐ด๐ ๐ฎ ๐บ๐ผ๐ฑ๐ฒ๐น ๐บ๐ถ๐๐๐ฒ๐. Grade findings, not just reports.
Still, SPEC-CXR is a gift to the field: a ready checklist and released code to do exactly that.
Disclosure: Iโm a co-author of GREEN and RadGraph-F1, two of the metrics this paper positions itself against.
Full breakdown in the first comment ๐
Blog post: jbdel.github.io/research-bloโฆ
Paper: papers.miccai.org/miccai-202โฆ
Code: github.com/lunit-io/spec-cxr
1
56
#MICCAI2025 HIGHLIGHTS #RadiologyAI ๐๐ต๐ฒ๐ฐ๐ธ๐ถ๐ป๐ด ๐ฎ๐ป ๐ซ-๐ฟ๐ฎ๐ ๐ฟ๐ฒ๐ฝ๐ผ๐ฟ๐ ๐ฎ๐ด๐ฎ๐ถ๐ป๐๐ ๐๐ต๐ฒ ๐ถ๐บ๐ฎ๐ด๐ฒ, ๐ป๐ผ๐ ๐ฎ ๐ฟ๐ฒ๐ณ๐ฒ๐ฟ๐ฒ๐ป๐ฐ๐ฒ. A paper reports 0.997 agreement with ground truth.
Blog post: jbdel.github.io/research-bloโฆ
Paper: papers.miccai.org/miccai-202โฆ
Data: huggingface.co/datasets/raziโฆ
17
#RadiologyAI at #NeurIPS2025
Interested in generating radiology reports? This work does the grunt, un-sexy work: modeling uncertainty.
The key insight is simple but powerful:
trustworthy AI doesnโt come from bigger models; it comes from explicitly modeling uncertainty and doing the hard reasoning work.
CURV's idea is to flip that:
explicitly label and ground uncertainty: whatโs unclear, where it is, and why.
1
265
#RadiologyAI at #neurips2025
OK, I love this one.
This is not, at first, radiology-related. But this one trick could absolutely translate to medical imaging. For architectures that don't use full self-attention, ๐ต๐ผ๐ ๐๐ผ๐ ๐ผ๐ฟ๐ฑ๐ฒ๐ฟ ๐ฝ๐ฎ๐๐ฐ๐ต๐ฒ๐ ๐ถ๐ป ๐๐ผ๐๐ฟ ๐ฉ๐ถ๐ง ๐ฎ๐ฐ๐๐๐ฎ๐น๐น๐ ๐บ๐ฎ๐๐๐ฒ๐ฟ๐.
Think about a chest x-ray, why should patches be read left-to-right, top-to-bottom like a PDF? Anatomy isn't linear. For chest X-rays, clinically meaningful context follows organs, symmetry, and spatial relations. If your ViT uses local or sparse attention, patch ordering becomes an inductive bias and we should design it to match anatomy.
Patch order โ implementation detail.
If you want to try this on open-source x-ray datasets, pm me ๐
2
211
#RadiologyAI at #neurips2025
So nice to see dinotxt has been applied to chest x-rays. Its actually an idea I wanted to try for a long time :-). Zero-shot segmentation is seriously impressive, and this paper shows you can precisely ground findings directly in the chest x-ray. Pretty neat...
5
249





