Here is a follow-up media news story on my project of organ and tumor segmentation platform for radiation therapy. The original article is in Chinese at https://mp.weixin.qq.com/s/vDzyKIKxuP7jCfmx6FPE4w.
I hope more doctors and physicists learn about NeuralRad, allowing the most advanced radiotherapy AI to benefit more patients and extend the lives of many.
I have decided to open user registration for use. Now, after signing up, you can start using this AI service immediately.
If you don't have the corresponding Dicom data for testing, you can click the Demo button on the platform, which will automatically open a case of a brain tumor MRI.
This is extremely cool, thanks for your work in this area, coolwulf. It's very important work and has the ability to help a lot of people. I work in an adjacent field and greatly appreciate seeing an open source approach to this problem.
This is amazing work! Can you divulge what institutions in Oklahoma are currently using NeuralRad? As someone in the state who often has MRIs looked at, this would definitely move a given medical provider higher on my list!
A naive question, maybe? I have about 5 years worth of CT, MRI, and PET data of my diagnosis (stage iv colon cancer) - to NED. I wouldn’t care about contributing it to any research if there was some way of anonymizing it. Or, does that kind of data automatically get contributed somehow? Some of the imaging was UCSF and some at Memorial Sloan-Kettering in New York.
Data would not get automatically contributed to research unless you explicitly gave informed consent for research studies. All researchers now require to maintain official records of informed consent given by patients or research participants. You could anonymize your data via DICOM anonymizer tools and offer them to UCSF or Memorial Sloan Kettering [or] you could host them on a website and offer to sign informed consent documents for researchers.
Speaking as a physician who works with both interventional and diagnostic radiology -- the self promotion and style of the post is offputting to say the least.
In addition, I'd never, ever, not in a million years, consider any interventional or diagnostic strategy from a blog post, most of which is written in a language, any language, I can not readily understand. Not to mention, I'd like to see names, not 1337 h4xx0r handles ("coolwulf" is great if you're 12 and playing WoW as a Worgen Deathknight, it's not cool if you'd like to convince me to add more tools to my workflow), when I consider modalities.
As a sidenote: modalities like this one are common in modern radio diagnostics. Harvard, MIT, Cambridge with King's, Paris Cité, and a few more are working on evidence based (and Open Source/Open Algorithm) approaches to AI diagnostics, all of which seem to have their ups and downs in outcomes. All their services are HIPAA compliant and certified as such, run in-house, do not require me to upload vast amounts of radiographic data to a website operated in a country I wouldn't trust with my daily egg consumption stats.
We're not talking funsies at the 7/11 here. We're talking diagnostic and therapeutic decisions. Basing those on a black box "pinky promise, it works" approach, is pretty much how people get killed. Not to mention, it's a 1980s view of things, the "if I can see it, I can cure it" approach. Modern oncology means to diagnose based on genetic and sequential markers, develop individualized strategies, long before imaging modalities become important. And if they do, there's plenty of hard- and software out there to make our lives so insanely easy, why send our stuff to a graphics card in China?
Also according to the article, those hospitals run the software HIPAA-compliant in-house, whereas the free online service is intended for people who literally couldn't afford anything else.
The resolution and planes in modalities that are useful for visual diagnostics (which, again, are outdated as hell) are so expensive, I don't think the little AI script drawing over suspected lesions (something a PET can do much better, including staging and grading) is the cost factor here.
But, hey, you do you. If you're comfortable subjecting yourself or your loved ones to this, by all means, do it. Without a decent sensitivity and specificity review, without an actual review in general, and without a corporation whose jugular I can cut if they screw up and murder my patient through a misdiagnosis, I'll stay away from someone's home grown shoe box medicine as far as I can. And I know all my serious colleagues will and are as well. It's not like that's the first person this week (or even today) trying to sell us one of those.
"Currently, many countries cannot operate Treatment Machines (radiation therapy machines) and CT scanners simultaneously due to insufficient power supply.
This might be hard to imagine for countries like China or the USA, but hospitals in these countries do not have hardware with sufficient computing power, nor do they have the funds to purchase AI service software. There's also a lack of understanding about AI technology, and even the performance of computers used by doctors can't be guaranteed."
By all means if you can afford it, sign an actual contract with the company developing this product; the free plan wasn't meant for you.
> "Currently, many countries cannot operate Treatment Machines (radiation therapy machines) and CT scanners simultaneously due to insufficient power supply.
Yes, I know. I worked in Ghana. And, know what? Unless you're running a $5m/month Cyberknife or similar, you don't do those dual modality approaches. Most, literally all except five or six research hospitals in the US and EU, treatments still work (very well) with lead marker lines on patients. We image, we look at the image we stage, we localize, we take out a tape measure, we draw. It might sound archaic, but it works extremely well, especially in places like Ghana.
I'd seriously love to see "coolwolf"s experience in developing country cancer treatments. I mean, in developing countries we deal 95% with cervix, breast, liver and prostate. Neither are hard to image and localize/stage. In the case of higher stages, exploratory imaging is also done, but those lesions aren't of initially surgical or radioherapeutic concern. Those who are, can be localized by eye only. And that's the ones, that software outlines.
I am talking in terms of my experience Treating multiple brain Mets patients. They have lesions as small as 0.01cc which we treat using either GammaKnife or CyberKnife with zero margin for CTV. This accuracy won't be achieved easily with tape measurements AFAIK.
It indicates, that he has sold a solution. It neither indicates that this solution is used[1], nor that those solutions are used for diagnostics.
And if Chinese single-individual solutions with gamer nicknames don't worry you, someone who frequents Hacker News and is probably not dumb, I understand many of the issues we have with medicine and medical communication much better. I'm sure you're confident that you, or a loved one, will be correctly diagnosed by this thing. I am not.
[1]: these hospitals are conglomerate hospitals who will buy things to try them out. I have dozens of bullshit solutions my bosses bought in storage. Why do you think you're getting raked over the coals for every small issue? Why do you think health care in the US is that expensive)
Thanks for your insights and sharing your expertise. I found it interesting to learn that these hospitals will purchase a solution just to test it out, possibly to then just shelve it. If that's what's happened here it seems quite dishonest of the author to imply that his software is being used there in actual clinical practice.
NeuralRad is being used during the clinic practice workflow at the moment. The platform currently is not FDA 510k cleared so we had to establish an IRB with the clinics which are building the platform into their clinical workflow.
You’re handling all the skepticism and hostility here with much more grace than I would be able to. I admire your tenacity and the scope of your effort-your hacker ethos, as it were.
I understand why the audience’s instinct is to judge you by the standards they’d judge a clinician or a formal medical device manufacturer, in that history of medical tech is littered with examples of well-meaning engineering efforts unintentionally causing harm.
As a counterbalance, though, I’d like to speak to the charitable interpretation: after all, how many times have I as a tech guy relied on practitioners in other fields to tell me what they could use, and whether what I built was helping them? It seems like you’re being judged on your skills as a practicing oncologist or full-scale US-market medical device manufacturer, when maybe a more fair frame might be that of a person who tries to help professionals whose work they admire by building tools they ask for.
I feel like just as it’s somebody else’s job to know how to doctor cancer, it will be somebody else’s job to prove that the tech is safe and appropriate to commercialize or popularize (what else is regulation for?).
From one person who likes making stuff to another person who makes stuff, though, I appreciate your good intentions, your creativity, and your follow through-and I admire your grace handling criticism here!
Welly he is working at the point where clinicians and medical device manufacturers meet. Hence, those are the only standards to measure against.
If those standards are met, great, more power to NeuralRad. It just rubs some people wrong to market this in classic SV start-up fashion, using the latest, in this case AI, hype. It just rubs people in more serious industrues the wrong way sometimes. Which, by the way, is valid feedback for everyone on HN with a B2B start-up targeting clients in very mature and risk averse indistries.
Yeah, that's where the serious business and online culture meet. And I know which one I prefer. That being said, with diagnostics becoming better each and every year, it is less straight forward to know which, if any treatment, to use against which tumor at a given time.
Generally, I do very much oreder if projects like this, regardless of the field, do include a healthy number of very experienced people in the actual domain, here that would be oncologists and radiologists specialized in oncology. Otherwise, I always assume it is some potentially dangerous almost-right first principle thinking at play. IMHO there is not a lot that is more dangerous than this.
Not so much business, as real medicine. The reason people get pissed at medicine is that it moves "slow" by their standards. But Jason M Somebody in his garage already built a cryo scalpel, why are you not using it?
Well, Thalidomide, the aforementioned cryo scalpel, Paolo Macchiarini, and others have taught us, that that's never a good idea.
It's 2024. We're throwing our lot in with serodiagnostics over anatomical localization, a PET scan being the only thing we need these days. If we know it's a lesion and we can FNA it, we will, else there's enough serodiagnostically we can do, to build individualized treatment plans (or determine if there's even a reason to use them).
It's pretty fun to see the computer draw little circles around lesions. The Bruker solution (also Open Source and free, by the by) does this admirably well. But it's neither useful in diagnostics nor in therapy, since both don't really hinge on a circle around a lesion.
But is an AI-generated circle! For sure there is a billion dollar start-up in there somewhere, right? /s
I see a lot of parallels between what you just wrote and big data and AI applications in a field I know a thing or two about: Logistics and Supply Chains. Same things happen there, a smart system highlighting a problematic order or time period. Great, but identifying those never was the real problem to begin with. At the same time, all the potential of using better planning tools to enable people or to automate repetitice tasks to free up peoples time get somewhat ignored because those use cases are less sexy than automating, and solving, all those "problems" using AI... But tjis hype cycle will ultimately be replaced with new one, like those before it. And I will yell at some other hype cloud!
Out of curiosity- the blog leans heavily on examples of clinicians operating in the developing world. Would the same lines of serodiagnostic techniques tend to be realistic in lower-resourced settings like that?
It seems plausible to argue that if you can afford a CT and a radiation therapy machine, then you probably meet the resource bar for the new family of techniques, but I just don’t have much of a reference point to judge.
> It seems plausible to argue that if you can afford a CT and a radiation therapy machine, then you probably meet the resource bar for the new family of techniques
Pretty much. The thing, though, is that places like Ghana (were I worked, I am a bit less educated about other places) get second hand machines from the US, Israel, UK, and Russia, so it'll take a few years until serodiagnostic and serotherapy labs "trickle down." So here he's right, not everyone has those resources.
However, the _imaging_ and _localization_ of lesions was never the issue.
Neither is the radiotherapeutic treatment. If you don't have enough power to do both at the same time, you won't have enough power to do his approach, since neither AI nor human eye can see through tissues. Humans move. Humans breathe. Either you take those movements as given, and live with the wider consequences, or you spend millions of cyber knives (https://med.stanford.edu/neurosurgery/divisions/radiosurgery...) that detect movement and counteract it. Having a more colorful version of the initial imaging won't change that. A radiologist's eye can do the same, and unless I see compelling evidence that the AI can do it better, something I have not seen, yet, any of the now 30+ solutions for AI analysis (as I said, Bruker and others are also offering theirs for free and open, and behind them stand massive research apparatuses) should especially not be used in developing countries.
Seriously, finding tumors and staging them is the easiest part of this job. And the rest can't be done by AI (yet.)
Thanks for your comment. However I don't think finding lesions is trivia especially if you are dealing with a patient with multiple brain Metastasis of more than 10 or even 20, 30 in counts.
For SRS, accurate localization and contour, label are crucial for the better outcome.
Thank you for your comment and I would like to clarify a few things:
1) NeuralRad could be running locally inside your own clinic. Currently, Stanford hospital, UTSW, Jefferson and several other clinics in the US are using it in house. In this configuration, the data never transferred outside the clinic.
2) The whole platform is HIPAA-complaint. All Dicom data are anonymized before submitting to NeuralRad server for inference if using the NeuralRad cloud server.
3) NeuralRad server is within US, not in "other" country.
At Stanford, the platform is used to help Physicians and Medical Physicists to contour Brain Mets. In addition, the platform never approves any dicom RT Structures. Physicians can use the generated RT structure to assist their SRS workflow.
I don’t know, I assume you are implying that you are on the hospital or SoM IRB or are privy to their work? I am curious to hear more of your opinions if you can share.
Would you please make your substantive points less aggressively? I appreciate that you know a great deal about this topic and we all probably have a lot to learn from you. Nonetheless you crossed the line here, and not only here—repeatedly mocking someone for their username, for example, is a cheap shot. You undermine your own arguments when you do things like this, in addition to poisoning the culture that we're trying so hard to prevent from becoming outright toxic.
Not quite. I am pointing out, that we're not talking about "yet another AI app" here, but something that might potentially be used to inform treatment decisions. In other words, something that decides over the life and death of a human being.
This life and death situation demands a certain decorum. Among others, to rise above the "hi I am coolwulf" stage of interpersonal interactions.
Let me go to a different topic in the same ballpark: imagine you had to build your SARS-CoV-2 response plan based on either a panel of women and men whose name and credentials you are privy to, or a single individual with a blog calling themselves "y0m0mm4" - which one would you, reasonably, choose to consider? Yes, the single individual may have, by chance or skill, happened upon the perfect plan. And, yes, the panel of experts may be completely wrong. But in medicine we call this a Zebra, a very rare event that, in a discipline of probabilistic empiricism, should not initially be considered.
I don't doubt, that all the above might come together to greatly improve treatment decisions. In fact, I know it, because as I said repeatedly, Bruker, Siemens, Tally, and others work on those solutions as well, and they are being used. Initially only in secondary verification, but with three FDA approvals expected this quarter, there might be much more coming.
I just don't think a pseudonymous account on Hacker News is the place to farm something that, not unlike many other medical interventions, has a massive potential of harm. As a somewhat related side note: the US was spared the horrors of thalidomide, because a single woman at the FDA (Frances Oldham Kelsey) refused to certify a medication that did not list the names of the inventors. When she did not get answers to her questions, largely because Grünenthal refused to let her speak to the drug's developers, she refused certification. In medicine names mean something, and if it's just the ability to verify background claims.
I don't think any professional is making treatment decisions based on random internet forum posts.
The purpose of HN is to have curious conversation about interesting things, and for that the high-order bit is that commenters be thoughtful and not aggressive.
As a medical physicist in radiation oncology who looks at segmentations generated like these I think you are missing the application slightly. The application of these segmentations is not in diagnosis or intervention, they are used mostly by a treatment planning algorithm.
The radiation oncologist will outline the tumour and a few organs at risk manually. This segmentation algorithm would then steps in and outlines organs that the doctor would not have traditionally contoured. For a lung lesion the RO may contour the lesion and the heart but might not contour both lungs and the diaphragm.
We can then input these segmentations into a treatment planning optimization algorithm that sets the radiation beam angles and collimation to meet constraints that minimize organ dose and maximize tumour dose. So in effect the application of this sort of segmentation is to give more information to an optimizer.
Not that it doesn't have its problems! But I think it's important to note that the application is not diagnostic
Really very cool, the time savings alone for places without autocontours is perhaps an hour per patient. Would be cool to see the following features in the next update - please dm if you want to collab!
1. Longitudinal comparison of images over time for response assessment
2. Prediction of radionecrosis vs. brain metastases
3. Flags for hypertrophy or atrophy of normal organs (i.e. cardiomegaly, splenomegaly) Easy to do based on volume
1. We have already developed the module for the longitudinal comparisons of the follow-up MR images. The module is not yet published but you can see the follow-up tab on the top of the dashboard (now grayed)
2. There's an on-going research between us and Stanford hospital on the prediction of radiation necrosis vs. brain mets. Our publication could be found in last year's AAPM conference.
Also I don't think you could DM on HN (maybe I'm missing something). But if your side wants to collaborate or implement this into your clinical workflow (We do provide free local installation of the whole platform inside your clinic network.), you could email me at coolwulf@gmail.com or hao.jiang@neuralrad.com
I hope more doctors and physicists learn about NeuralRad, allowing the most advanced radiotherapy AI to benefit more patients and extend the lives of many.
I have decided to open user registration for use. Now, after signing up, you can start using this AI service immediately.
If you don't have the corresponding Dicom data for testing, you can click the Demo button on the platform, which will automatically open a case of a brain tumor MRI.