JB (@IAMJBDEL) profile picture

Stanford - Radiology AI | RadAI @ HOPPR | Previous: ML @HuggingFace, Academic staff - Research @Stanford University, @StanfordAIMI Affiliate

Palo Alto
Joined July 2017
Photo by JB (@IAMJBDEL)Photo by JB (@IAMJBDEL)Photo by JB (@IAMJBDEL)Photo by JB (@IAMJBDEL)Photo by JB (@IAMJBDEL)Photo by JB (@IAMJBDEL)Photo by JB (@IAMJBDEL)Photo by JB (@IAMJBDEL)Photo by JB (@IAMJBDEL)
#RadiologyAI Our report drafting solution is deployed in 300 sites over 3 countries. Here are ๐˜๐—ต๐—ฒ ๐—ผ๐—ฝ๐—ฒ๐—ป ๐—พ๐˜‚๐—ฒ๐˜€๐˜๐—ถ๐—ผ๐—ป๐˜€ 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 ๐˜‚๐—ป๐—ฝ๐—ฟ๐—ฒ๐—ฐ๐—ฒ๐—ฑ๐—ฒ๐—ป๐˜๐—ฒ๐—ฑ ๐— ๐—Ÿ ๐—ฐ๐—ต๐—ฎ๐—น๐—น๐—ฒ๐—ป๐—ด๐—ฒ๐˜€: 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. ๐—ง๐—ต๐—ฒ ๐—บ๐—ผ๐—ฑ๐—ฒ๐—น ๐—บ๐˜‚๐˜€๐˜ ๐—ฎ๐—ฑ๐—ฎ๐—ฝ๐˜ ๐˜๐—ผ ๐˜๐—ต๐—ฒ๐—บ, ๐—ป๐—ผ๐˜ ๐˜๐—ต๐—ฒ ๐—ผ๐˜๐—ต๐—ฒ๐—ฟ ๐˜„๐—ฎ๐˜† ๐—ฎ๐—ฟ๐—ผ๐˜‚๐—ป๐—ฑ. ๐—”๐—ป๐—ฑ ๐˜๐—ต๐—ฎ๐˜ ๐—ถ๐˜€ ๐—ฝ๐—ฒ๐—ฟ๐—ณ๐—ฒ๐—ฐ๐˜๐—น๐˜† ๐—ณ๐—ถ๐—ป๐—ฒ. 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 ๐—บ๐—ผ๐—ฟ๐—ฒ ๐˜๐—ต๐—ฎ๐—ป ๐Ÿฎ,๐Ÿด๐Ÿฌ๐Ÿฌ ๐—ถ๐—บ๐—ฎ๐—ด๐—ถ๐—ป๐—ด ๐—ฑ๐—ฒ๐—ฝ๐—ฎ๐—ฟ๐˜๐—บ๐—ฒ๐—ป๐˜๐˜€ ๐—ถ๐—ป ๐Ÿฎ๐Ÿฌ ๐—ฐ๐—ผ๐˜‚๐—ป๐˜๐—ฟ๐—ถ๐—ฒ๐˜€, giving us a unique opportunity to solve these challenges at scale.
1
6
699
#RadiologyAI New proceedings alert ๐Ÿšจ . Finally, someone does the painstaking groundwork. ๐—ก๐—œ๐—›โ€™๐˜€ ๐—–๐—ง-๐—•๐—ฒ๐—ป๐—ฐ๐—ต ๐—ฝ๐˜‚๐˜๐˜€ ๐˜„๐—ผ๐—ฟ๐—ฑ๐˜€ ๐—ผ๐—ป ๐Ÿฎ๐Ÿฌ,๐Ÿฏ๐Ÿฏ๐Ÿฑ ๐—–๐—ง ๐—น๐—ฒ๐˜€๐—ถ๐—ผ๐—ป๐˜€ (๐— ๐—˜๐—Ÿ๐—•๐—” ๐Ÿฎ๐Ÿฌ๐Ÿฎ๐Ÿฒ). ๐—ข๐—ป ๐—ถ๐˜๐˜€ ๐—ณ๐—ผ๐˜‚๐—ฟ-๐—ฐ๐—ต๐—ผ๐—ถ๐—ฐ๐—ฒ ๐—ฒ๐˜…๐—ฎ๐—บ, ๐—š๐—ฃ๐—ง-๐Ÿฐ๐—ฉ ๐˜€๐—ฐ๐—ผ๐—ฟ๐—ฒ๐˜€ ๐Ÿฎ๐Ÿด%. 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 ๐Ÿฎ๐Ÿด%. It is a 2023โ€“2024 model; nothing newer was tested. ๐—”๐—ป๐—ฑ ๐—ฟ๐—ฎ๐—ฑ๐—ถ๐—ผ๐—น๐—ผ๐—ด๐—ถ๐˜€๐˜๐˜€? ๐—ง๐˜„๐—ผ ๐˜€๐—ฒ๐—ป๐—ถ๐—ผ๐—ฟ ๐—ฟ๐—ฎ๐—ฑ๐—ถ๐—ผ๐—น๐—ผ๐—ด๐—ถ๐˜€๐˜๐˜€ ๐—ด๐—ผ๐˜ ๐—ฎ๐—ฏ๐—ผ๐˜‚๐˜ ๐Ÿด ๐—ถ๐—ป ๐Ÿญ๐Ÿฌ ๐—ฟ๐—ถ๐—ด๐—ต๐˜, ๐—ผ๐—ป ๐Ÿญ๐Ÿฌ๐Ÿฌ ๐—ฐ๐—ฎ๐˜€๐—ฒ๐˜€. โ†’ The best AI in the paper, a model trained on the released lesions, gets ๐Ÿฑ๐Ÿฌ% 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. ๐—ช๐—ต๐—ฎ๐˜ ๐˜†๐—ผ๐˜‚ ๐—ฐ๐—ฎ๐—ป ๐˜‚๐˜€๐—ฒ ๐˜๐—ผ๐—ฑ๐—ฎ๐˜†: โ†’ ๐Ÿฎ๐Ÿฌ,๐Ÿฏ๐Ÿฏ๐Ÿฑ ๐—–๐—ง ๐—น๐—ฒ๐˜€๐—ถ๐—ผ๐—ป๐˜€, each with a one-sentence description, a size and a box, plus 2,850 exam questions with the paperโ€™s scores to compare against. On Kaggle (link in the first comment). โ†’ The licence is CC BY-NC-SA 4.0: free for research, ๐—ป๐—ผ ๐—ฐ๐—ผ๐—บ๐—บ๐—ฒ๐—ฟ๐—ฐ๐—ถ๐—ฎ๐—น ๐˜‚๐˜€๐—ฒ. The lesson: before trusting a model with a whole report, test it lesion by lesion: ๐˜„๐—ต๐—ฒ๐—ฟ๐—ฒ ๐—ถ๐˜ ๐—ถ๐˜€, ๐˜„๐—ต๐—ฎ๐˜ ๐—ถ๐˜ ๐—ถ๐˜€, ๐—ต๐—ผ๐˜„ ๐—ฏ๐—ถ๐—ด ๐—ถ๐˜ ๐—ถ๐˜€. 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 ๐Ÿ‘‡
3
4
13
1,481
#RadiologyAI My highlight of #MICCAI2026 ๐—Ÿ๐˜‚๐—ป๐—ถ๐˜โ€™๐˜€ ๐—ฆ๐—ฃ๐—˜๐—–-๐—–๐—ซ๐—ฅ ๐—ด๐—ฟ๐—ฎ๐—ฑ๐—ฒ๐˜€ ๐—”๐—œ ๐—ซ-๐—ฟ๐—ฎ๐˜† ๐—ฟ๐—ฒ๐—ฝ๐—ผ๐—ฟ๐˜๐˜€ ๐—ณ๐—ถ๐—ป๐—ฑ๐—ถ๐—ป๐—ด ๐—ฏ๐˜† ๐—ณ๐—ถ๐—ป๐—ฑ๐—ถ๐—ป๐—ด. ๐—ง๐—ต๐—ฒ ๐—ฐ๐—ฎ๐˜๐—ฐ๐—ต: ๐—ผ๐—ป ๐—บ๐—ผ๐˜€๐˜ ๐—ธ๐—ถ๐—ป๐—ฑ๐˜€ ๐—ผ๐—ณ ๐—ณ๐—ถ๐—ป๐—ฑ๐—ถ๐—ป๐—ด, ๐˜๐—ต๐—ฒ ๐—ฏ๐—ฒ๐˜€๐˜ ๐—”๐—œ ๐—ถ๐˜€ ๐—บ๐—ผ๐˜€๐˜๐—น๐˜† ๐˜„๐—ฟ๐—ผ๐—ป๐—ด. 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 ๐Ÿต ๐—ผ๐—ณ ๐Ÿญ๐Ÿฑ 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. ๐—ช๐—ต๐—ฎ๐˜ ๐˜†๐—ผ๐˜‚ ๐—ฐ๐—ฎ๐—ป ๐˜‚๐˜€๐—ฒ ๐˜๐—ผ๐—ฑ๐—ฎ๐˜†: 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 paperโ€™s eight judges, agreement with radiologists ranges from ๐Ÿฌ.๐Ÿฒ๐Ÿฌ ๐˜๐—ผ ๐Ÿฌ.๐Ÿณ๐Ÿด. 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: ๐—ผ๐—ป๐—ฒ ๐˜€๐—ฐ๐—ผ๐—ฟ๐—ฒ ๐—ฝ๐—ฒ๐—ฟ ๐—ฟ๐—ฒ๐—ฝ๐—ผ๐—ฟ๐˜ ๐—ฐ๐—ฎ๐—ปโ€™๐˜ ๐˜€๐—ฎ๐˜† ๐˜„๐—ต๐—ถ๐—ฐ๐—ต ๐—ณ๐—ถ๐—ป๐—ฑ๐—ถ๐—ป๐—ด๐˜€ ๐—ฎ ๐—บ๐—ผ๐—ฑ๐—ฒ๐—น ๐—บ๐—ถ๐˜€๐˜€๐—ฒ๐˜€. 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: Iโ€™m a co-author of GREEN and RadGraph-F1, two of the metrics this paper positions itself against. Full breakdown in the first comment ๐Ÿ‘‡
1
2
4
463
#MICCAI2025 HIGHLIGHTS #RadiologyAI ๐—–๐—ต๐—ฒ๐—ฐ๐—ธ๐—ถ๐—ป๐—ด ๐—ฎ๐—ป ๐—ซ-๐—ฟ๐—ฎ๐˜† ๐—ฟ๐—ฒ๐—ฝ๐—ผ๐—ฟ๐˜ ๐—ฎ๐—ด๐—ฎ๐—ถ๐—ป๐˜€๐˜ ๐˜๐—ต๐—ฒ ๐—ถ๐—บ๐—ฎ๐—ด๐—ฒ, ๐—ป๐—ผ๐˜ ๐—ฎ ๐—ฟ๐—ฒ๐—ณ๐—ฒ๐—ฟ๐—ฒ๐—ป๐—ฐ๐—ฒ. A paper reports 0.997 agreement with ground truth.
1
1
2
517
โ€ข No code or model listed; omitted findings are out of scope. Great work by Razi Mahmood and co-authors (RPI, IBM Research, Stanford, MGH)
1
14
#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 doesnโ€™t 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: whatโ€™s 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, ๐—ต๐—ผ๐˜„ ๐˜†๐—ผ๐˜‚ ๐—ผ๐—ฟ๐—ฑ๐—ฒ๐—ฟ ๐—ฝ๐—ฎ๐˜๐—ฐ๐—ต๐—ฒ๐˜€ ๐—ถ๐—ป ๐˜†๐—ผ๐˜‚๐—ฟ ๐—ฉ๐—ถ๐—ง ๐—ฎ๐—ฐ๐˜๐˜‚๐—ฎ๐—น๐—น๐˜† ๐—บ๐—ฎ๐˜๐˜๐—ฒ๐—ฟ๐˜€. 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 ๐Ÿ˜‰
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