Daniel Donoho, MD @ddonoho

Calling all residents and medical students who want to join us in surgical AI research with partners at @MayoClinic @StanfordNsurg @NeurosurgUCSF and more!

Surgical Data Science Collective @SurgicalDSC

4 weeks ago

🌟 Are you a medical student or resident that is curious about the future of AI in surgery? On Monday, October 5th at 2pm Eastern, SDSC is hosting an introductory session for those interested in surgical innovation, AI, research, and the growing role of surgical video in

0 0 6 1K 5

🗣️ Last week, we brought together surgeons, computer scientists, and researchers in NY to discuss a deceptively difficult question: how do we know when AI actually understands surgery? 💡 The evening sparked a lot of thoughtful discussion about where surgical AI is today, and what it will take to move the field forward. One thing is clear: better surgical AI starts with better surgical data. Our Founder, @ddonoho , opened with the clinical motivation behind SDSC’s work and why understanding the context of surgery matters. Our CTO @genuinebombayit , and Senior ML Engineer Jack Cook provided an in-depth look at how we’re teaching AI to identify patterns in surgical video and distinguish what the surgeon is doing from what the camera happens to see. 🩺 The goal is bigger than recognizing what happens in an operation. It’s understanding the patterns of surgery well enough to better quantify surgical processes and decision-making, and evaluating how specialized surgical models compare with today’s frontier AI models. 🌟 We were also joined by @XYHan_ and Dr. Kee Park, who helped us widen the conversation beyond what's possible today and the challenges we still need to solve to move surgical data science forward. Every surgical video contains information about technique, decisions, anatomy, instruments, actions, and outcomes. Bringing more of that information together across surgeons, institutions, procedures, and geographies gives researchers the opportunity to build and evaluate AI on a much broader picture of surgery. 👉 That’s why we’re building the Surgical Video Platform (SVP). If you or your institution record surgical video, we’d love to work with you. Contributing de-identified video can help create the diverse evidence base needed to develop, test, and improve the next generation of surgical AI. Try it for yourself: surgicalvideo.io/surgical-video… And if our work sparked an idea for a research collaboration or another way we could work together, we’d love to hear from you: surgicalvideo.io/research 👐 Thank you to everyone who joined us in New York, shared their perspective, and helped push the conversation forward. #SurgicalAI #SurgicalTraining #SurgicalEquity

keyboard_arrow_left Previous keyboard_arrow_right Next

What if AI could actually understand what happens during surgery? Not simply recognize an image - but identify surgical actions, track instruments, assess technical skill, recognize critical steps, and potentially predict whether a procedure will succeed. That is the emerging world of surgical data science. On this episode of @ProgressPotent1, I’m joined by Dr. Daniel Donoho, MD - @ddonoho - Neurosurgeon at @childrensnatl and @gwsmhs, Founder of the @surgicaldsc and Co-Founder of the Foundation for Digital Neurosurgery. Dr. Donoho's work sits at a fascinating intersection of neurosurgery, computer vision, machine learning and surgical education. We discuss: 🧠 How AI can decode what surgeons are doing from video; 🤖 Whether AI can objectively measure surgical skill; 🩸 How algorithms can predict hemorrhage outcomes from surgical video; 🎥 Why millions of hours of surgical footage could represent an enormous untapped dataset; 🔬 Creating a standardized “language” for surgical actions and gestures; 🏥 Whether AI could provide continuous feedback inside the operating room; 🌎 How surgical data could democratize access to expertise around the world; 🧠 And where all of this may ultimately lead as brain-computer interfaces and neurotechnology begin generating entirely new forms of human neurological data; The big question: Are we moving from recording surgery to actually teaching machines to understand it? And if the brain itself becomes a data-generating interface… Who owns that data? #Neurosurgery #SurgicalAI #ArtificialIntelligence #SurgicalDataScience #DigitalSurgery #Neurotechnology #BCI #MedicalAI #FutureOfMedicine #ProgressPotentialAndPossibilities @ampragency

We've updated the most complete surgical intelligence leaderboard of current Vision Language Models (VLMs), benchmarking them on 10 datasets: eval.surgicalvideo.io All VLMs are significantly below the specialized model ceiling More in our paper: arxiv.org/abs/2603.27341

keyboard_arrow_left Previous keyboard_arrow_right Next

something non-obvious about our surgery benchmark: it doesn't look like the math benchmarks. not all specialized skills (surgery vs math) are created equal. you'd think that they'd be similar given both surgery and math require a lot of specialized knowledge possessed by a limited group of specially-trained people. in the math benchmarks, Astra/Fable moggs and then everything else is just foundation models in descending order of released data/size. fine-tuned mini-models are non-players on frontiermath, etc. whereas here fine-tuned mini-models (and not-so-small but still fine-tuned models like LemonFM) somehow get to the top. these benchmarks were created by other independent teams so it's not like we're tipping the scales in our favor somehow (i promise we arent) besides, of course, fine-tuning which is the point, anyways. the reason is that the entirety of math is "publishable" (the proofs, steps, problem) whereas here surgery is basically a surgeon acting on instinct/experience that is not entirely write-down-able from their med school/residency/fellowship training. in fact, you can go to two different hospitals and the procedure for the *same surgery* will look entirely different up to some high level goals and steps. plus, you have to be careful with medical data because of HIPAA etc which leads to overheads and limitations due to anonymizations etc. so not all specialized skills are created equal. the ones that can be more or less written down in their entirety (math) will have lots of data and will be internalized by GPT/Claude pretty quickly. the ones that are based on "experience training" (like surgery) aren't so amenable to the current models. and, no, you can't just make the llm read a medical textbook and assume it generalizes (we tried that in the paper).

Kirill @kskblv

a month ago

We've updated the most complete surgical intelligence leaderboard of current Vision Language Models (VLMs), benchmarking them on 10 datasets: eval.surgicalvideo.io All VLMs are significantly below the specialized model ceiling More in our paper: arxiv.org/abs/2603.27341

keyboard_arrow_left Previous keyboard_arrow_right Next

0 2 5 1K 0

🌟 SDSC founder @ddonoho recently joined @FRANCE24 to talk about our work with surgeons in Tanzania and the role AI can play in expanding surgical capacity. One of Dr. Donoho's points gets right to the heart of why we’re there: Tanzania already has exceptional surgeons doing incredibly complex work, so how do we take the excellence that exists in a few hands and spread it as broadly as possible? 🖥️ Surgical video is part of the answer. Through our work with colleagues in Tanzania and Children's National Hospital, we’re looking at how AI can help surgeons learn from existing expertise, including in particularly challenging neurosurgical procedures. 🌍 Dr. Donoho also touches on something that matters enormously as surgical AI develops: local data matters. A model built on generalized surgical knowledge still needs to be tested with local hospitals, surgeons, procedures, and data to know whether it actually works in that setting. Our work in Tanzania grew out of years of relationships, including Dan’s friendship with Dr. Hamisi Shabani and their efforts to bring together the country’s neurosurgical community. Those relationships ultimately helped lead to SDSC’s first program in Tanzania at Muhimbili Orthopaedic Institute. 🩺 The technology may be new. The goal is much more fundamental: help more surgeons learn from great surgeons, so more patients can access great surgical care. 📺 Watch the France 24 segment (timestamp at 8:55): france24.com/en/kenya-crack… #SurgicalAI #GlobalSurgery #Neurosurgery #GlobalHealthEquity

We benchmarked GPT-6 Astra on the @SurgicalDSC @CAAI_Booth Surgical AI Leaderboard - astra is the new generalist leader, but falls behind tiny specialist models (1000x smaller) Full Leaderboard: skblv.github.io/neurosurgery-v… Paper: arxiv.org/abs/2603.27341 HT @kskblv @EricFithian

keyboard_arrow_left Previous keyboard_arrow_right Next

Kirill @kskblv

2 months ago

We benchmarked Claude Fable 5.1 and Gemini 3.8 Flash on surgical intelligence: - The performance of the two new releases is spiky (e.g. good on tools, bad on VQA) - Frontier LLMs still underperform tiny specialized models Check out the full LLM leaderboard linked below

2 4 15 7K 7

Work hard, play hard 💪 We host a department dinner during each visiting speaker's time on campus. Excited to learn more about the research endeavors of Dr. Dan Donoho, a data driven surgeon scientist working hard to provide access across boarders! #neurosurgery #globalhealth

keyboard_arrow_left Previous keyboard_arrow_right Next

Neurosurgery Grand Rounds (@NeurosurgeryOSU) today with Dan Donoho, M.D., (@ddonoho) of Children’s National Hospital (@ChildrensNatl) and George Washington University (@GWtweets). He showed us how AI and computer vision are changing neurosurgery: teaching machines to understand what’s happening in the OR, flagging the moments that drive outcomes, and turning surgical video into a training tool for residents. The goal isn’t smarter software. It’s safer operations and better surgeons. @OSUWexMed @OhioStateMed #AI #neurosurgery #surgery #training

keyboard_arrow_left Previous keyboard_arrow_right Next

@CNS_Update @AANSNeuro folks: strongly recommend attending this upcoming great talk!

Surgical Data Science Collective @SurgicalDSC

2 months ago

🧠 Throughout #NeurosurgeryAwarenessMonth, we've been exploring how AI can help make aspects of neurosurgical care more measurable. Our next Data Science Roundtable brings that conversation into the neurocritical care setting. 📣 Join Dr. Rui Feng for a discussion on

0 0 1 509 0

1/ 🧠 Recognizing what's in a surgical video is one thing. Understanding what's happening is another. 👩🏽‍🏫 Before AI can understand surgery, we first have to teach it the language. When SDSC trains a surgical tool detection model, thousands of human-verified annotations help teach the model what it's seeing. A scalpel needs to be a scalpel. A curette needs to be a curette. Actions, anatomy, phases, and eventually more complex surgical concepts all require clearly defined ground truth. 📢 We’ve all heard the adage "Garbage in, garbage out." A mislabeled example can become a mistake the model learns. That's why human oversight isn't something surgical AI eventually graduates from. Humans define the ontology, verify the data, and ultimately determine whether a model's output has meaning in the real world. Getting the labels right is only the foundation; just because a model recognizes a surgical instrument in a single frame does not mean it understands an operation. 🩺 Surgery unfolds over time. Actions combine into maneuvers. Maneuvers form tasks. Tasks become procedural steps. Together, they reveal the structure of an operation. 🤖 Teaching AI to represent those relationships is part of the transition SDSC describes as Digital Surgery 1.0 → Digital Surgery 2.0. Digital Surgery 1.0 gave us cameras, robotic systems, tracking tools, and other technologies capable of digitizing and sensing the operating room. Digital Surgery 2.0 asks what happens when we can begin to understand, measure, and learn from the surgical processes those technologies capture. For #NeurosurgeryAwarenessMonth, that's the future we're working toward at SDSC: not AI that simply recognizes what's visible in neurosurgical video, but models that can help us understand how surgery unfolds and ultimately turn that understanding into objective, clinically meaningful surgical metrics. And it starts with something decidedly human: teaching the model what matters.

keyboard_arrow_left Previous keyboard_arrow_right Next

@nikillinit @ashokaraj_ could lower efficacy be associated with higher adherence to... eg if I lose those 15 lbs in 3 months, I'll drop out but if it takes me 9 months, I'll stick for the 9 months ...

@KMeierNSICU @claudeai love this. everyone should. have one such thing/idea ... and if there are common tasks (with frameworks!) and benchmarks, so much the better

Physician, benchmark thyself! [solve for the equilibrium]

Isaac Kohane @zakkohane

3 months ago

Two years ago I took two hundred pairs of simulated patients and did something boring: I decided, one pair at a time, which of two patients would be seen first. I asked state-of-the-art LLMs (@google @openai @AnthropicAI) to do the same with the same patients. I just repeated

5 6 39 8K 16

Fable 5 and GPT-5.6-Sol vs finetuned YOLO on surgical benchmarks. Takeaway here is that the tiny finetuned YOLOv2 (with ~26M params) still beats the SOTA out-off-the-shelf models. Specialized training is still needed for specialized tasked. Credit: @kskblv

Kirill @kskblv

5 months ago

We benchmarked Fable on surgical tool detection so you don’t have to. At almost 2 cents/image, Claude Fable 5 still loses to a compact image classifier. Will frontier VLMs ever catch up in surgical perception? Read our paper to learn more: arxiv.org/abs/2603.27341

0 4 6 4K 1

@kskblv Related paper. Collab with @SurgicalDSC : arxiv.org/abs/2603.27341

@AmandaAskell Have you tried asking ML scientists “how much data do you need to do <<x>> “ ;)

@nikillinit Agree ... we just require an actual benchmark that is live and re-runnable ... no more static papers.

Congratulations to the R Project on receiving the $1M Rousseeuw Prize for Statistics 2026! @RProgramming @RConsortium @Rbloggers We now take for granted computational science requires a strong ecosystem of code sharing. However, before uv, conda, and pip, it was R that laid the foundations for the dissemination of statistical/computational methodologies not only to statisticians but also to biologists, social scientists, and other disciplines driven by data analysis. We now rest under the shade of large trees, but it is gratifying to see the Rousseeuw Prize go to the R Project which was one of the first plant the seeds and till the soil --- and kept it fertiles for almost three decades. prweb.com/releases/2026-…

@RussellShaw_MLS We have a solution for this @ChildrensNatl … MRI at the point of care. Rapid, effective, kid friendly.