Hacker News new | past | comments | ask | show | jobs | submit login
AI can detect coronavirus from CT scans in twenty seconds (linkedin.com)
109 points by laurex on March 13, 2020 | hide | past | favorite | 73 comments



So many things.

"It takes doctors around 20 mins to do the same" is patently false. I would say a full review and dictation of a chest CT by a trained radiologist takes... 10 minutes? 20 if really complicated (metastatic cancer). But for the radiologist to decide whether or not there were findings compatible with COVID-19 (i.e. what this algorithm is performing) would probably take seconds.

In fact, you could probably train almost any adult (and probably most children!) to distinguish between "possibly COVID" and "no COVID" lung CTs very quickly.

However, as pointed out elsewhere, the presence of these findings (such as what we call "ground glass opacities" -- which is just a fancy/weird way to describe areas that are slightly less transparent than normal lung but not quite opaque) is not specific at all. They can be caused by COVID-19 or any number of other viral or fungal lung infections, as well as by reactions to drugs, hemorrhage or edema in the lungs, and so on.

A positive CT would almost certainly require a confirmatory COVID-19 test. A negative CT scan in a patient with symptoms... would still demand a COVID-19 test, because findings in the lungs often lag behind the infection.

So CT doesn't change your decision-making in the screening process. And they are time-consuming, costly, and harmful (they use radiation!). Their utility in caring for COVID-19 patients is probably to assess severity of disease, monitor response, and to rule out additional pathologies in the lungs. (Not to say AI couldn't help with those!)


It also doesn't sound like it even lowers the bar for getting a diagnosis, even with all these caveats. You'd still need a CT scan, in a world where many can't even get a respirator.

I'd be more interested in a model that tries to give a preliminary diagnosis (positive or negative) based on data a patient already has with them or can get at home. An image of their throat, audio of their cough, maybe a time sequence of thermometer readings. If we could self-diagnose (or at least rule out some cases) using a smartphone, that would truly change the game.


I think everything that you've said is completely true. However, I'll add that it matters a lot more if you otherwise don't have testing capacity available, or if your test has a large false negative rate. China relied heavily on CT-based diagnosis to backstop a lack of available and reliable testing.

Another issue is that modern sanitation standards don't really enable using a CT to diagnose an infectious disease. The equipment is far too slow and expensive to truly sanitize, you would have to be willing to just PPE the patient and hope that works out. Again, perhaps better than skipping any diagnostic testing altogether if you don't have any alternative, but certainly not a first-line choice, and ethically questionable if you risk exposing uninfected people.


Actually you just need to have them wear a mask and wipe down the equipment with standard hospital grade wipes. We know now that COVID-19 is not airborne and is spread by droplets only so this would be routine precautions, unlike TB which we scan routinely and is an airborne pathogen.

The problem with CT is we know that the test is completely normal in 50% of patients from day 0-2, and 10-15% of patients up to day 5, and roughly 7-10% of patients up to day 10. These numbers are probably an under representation given how few people with the illness get CT imaging.

Additionally as has been stated the findings are considered only somewhat specific, we believe that specificity will decrease as more and more people will eventually get scanned and prove our findings thus far to be false.

That is why we (in radiology) currently recommend against the routine use of CT in screening or diagnosis of COVID-19 related illness as a negative test does not mean anything and a positive test is usually noncontributory (clinical diagnosis of COVID-19 infection already made). Additionally, this will rarely affect clinical management. The only potential argument being made is the small subset of people who have viral pneumonia and may theoretically avoid being asked to self-quarantine however I highly doubt in the present climate any infectious patient with a cough will be allowed/think it wise to roam the streets.

China has also removed CT from their sixth edition diagnostic manual for this disease as an aside.

Source is from STR/ACR, there is a good video below: https://veritastv.org/programs/covid-19


> We know now that COVID-19 is not airborne and is spread by droplets

What?! It was only, like, yesterday when I heard that "we know" pretty much the opposite. I'm less and less convinced "we know" anything at all.


Most nations have follow the WHO advice and have all dropped the requirement for airborne precautions/isolation rooms and we now use droplet precautions as we do with every other influenza and Coronaviruses. It is my understanding that the CDC has done so as well, partially due to supply chain issues but also largely given the lack of evidence of airborne transmission.

https://www.publichealthontario.ca/-/media/documents/ncov/up...

https://apps.who.int/iris/bitstream/handle/10665/331215/WHO-...


There is a medRxiv paper from March 9 that suggests 3 hours half life in the air

Anyone who “knows” something is currently wrong, but there is evidence of possible airborne transmission


Can you provide this link?

I could only find:

https://www.medrxiv.org/content/10.1101/2020.03.09.20033217v...

Which does not actually address whether there is airborne transmission of the virus but rather refers to laboratory generated aerosolized particles and viability on various surfaces.

If you know otherwise please share as this is in direct contradiction to what various health organizations have been telling HCWs and forms the basis of our hospital’s standard of care which like all other in North America is that airborne precautions are NOT needed except during aerosolized particle generating procedures.


“Standard of care” is risk managed by a vast bureaucracy of counter-balanced interests. It is not designed to incorporate rapid learning of emerging unknowns.

During the exponential introduction of a new disease, the standard of care lags, badly.

Repeating “that goes against the basis” or “not needed” doesn’t make sense until the competing interests have settled on a new standard incorporating understanding of the new unknown.

Better to follow the emerging research, tailor care, document, and contribute findings back, to optimize the learning curve.


I don’t follow your logic. When COVID-2019 first became a thing the “standard of care” was to use airborne precautions and respirators, it did not “lag badly” as you so claim.

Two months of research later there has been no evidence that this virus can be transmitted by airborne particles (read: not droplets) and the standard of care was updated a couple of days ago to reflect that, hence why we now do not require negative pressure isolation rooms and respirators.

I don’t see where the lag is.


Here's yesterday's updated standard of care language in regional guidance. Recall that a standard of care also protects providers and facilities from liability. By relaxing it, both are less liable for things that may still be risks.

Looks like the language makes it clear the goal here is to preserve the supply of masks:

The safety of healthcare workers is a top priority for the Health Department. As we gain more understanding of COVID-19, our guidance will evolve. The use of standard, contact, and droplet precautions with eye protection is appropriate when caring for patients with possible or confirmed COVID-19. Personal protective equipment (PPE) should include: facemask (procedure or surgical mask) AND gown AND gloves AND eye protection (goggles or face shield).

The Health Department recommends healthcare workers do not need to use a fit tested N95 respirator or Powered Air Purifying Respirator (PAPR) for routine (non-aerosol generating) care of a COVID-19 patient. Patients can be evaluated in a private examination room with the door closed.

Supplies of PPE must be reserved for high risk procedures due to potential supply chain constraints.

Ample studies indicate the safety of droplet precautions which may also help prevent the complete exhaustion of fit tested N95 respirators and PAPRs; higher level PPE will continue to be needed to protect HCWs during critical and medically necessary aerosol generating procedures (e.g., intubation, suctioning) throughout the course of this outbreak.

Placing the patient in an AIIR and the use of a fit tested N95 respirator or PAPR is still recommended for aerosol-generating procedures (e.g., intubation, suctioning, nebulizer therapy) and when caring for patients with severe illness requiring intensive care.


Previously, you wrote, "various health organizations have been telling HCWs and forms the basis of our hospital’s standard of care which like all other in North America is that airborne precautions are NOT needed".

Now you write, "'standard of care' was ... two months of research later ... the standard of care was updated a couple of days ago to reflect that".

Yet you comment, "I don’t see where the lag is."

The two months' lag was enough it's not at all obvious which standard of care you were referring to when you're saying no lag, the lagged standard or the just updated standard.

This very comment thread highlights the issue.


The article fails to ask the obvious next question: if this can potentially apply to much cheaper and instantaneous tests, maybe something ultrasound based? I'm not an expert on medical imagining, so can someone comment on ultrasounds' usefulness here?


Ultrasound has a ton of untapped utility and is a fantastic medical imaging modality. Like you said, the equipment is cheap and portable (can be pocket sized, even!). It’s near instantaneous, allowing you to image in real time. You can even measure blood flow velocity and see other physiology. (I could go on!)

But its great advantage over x-ray and CT, which is that it doesn’t use ionizing radiation, is also its greatest weakness. Ultrasound uses sound waves, of course, which unfortunately simply cannot travel very far in the body, and (critically) do not travel well across certain interfaces. One such interface is with air. Sound waves travel through your chest wall fine but on arriving at the lung... they just sort of fizzle out. Very little sound continues, and even less comes back to the transducer as meaningful signal.

So ultrasound is unfortunately very poor at seeing lung tissue. (It can show us some secondary findings of pulmonary disease in the lungs, such as fluid collecting in the thoracic cavity [a pleural effusion], but that’s yet another non-specific finding.)


Not entirely true.

Ultrasound actually has pretty good sensitivity for pneumonia (88%) and even higher if you look at the “difficult” cases requiring CT in the Cochrane review I link to below.

While the fundamentals of what you say are correct, consolidation by definition is the build up of fluid which is very well depicted by ultrasound. For a predominately peripheral airspace disease like COVID-19 ultrasound should theoretically be very sensitive.

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5279077/


I wasn’t aware lung US was at this point — I’ll have to read more!


> And they are time-consuming, costly, and harmful (they use radiation!)

To put things in perspective, I got a CT in China, and the imaging took less than 5 minutes, and cost $35 (just the machine, not reading). The radiation is less than the increased radiation than taking a flight to China.

While your perspective is probably accurate for the US, if you step back and think, it is no more complicated than getting grade school photos.


By my math, a chest CT is about 10x the dose from a flight. However, there is no “safe” level of ionizing radiation. There is only acceptable and unacceptable risk.


Right, one CT scan increases your lifetime risk of cancer 0.1%: https://www.scientificamerican.com/article/how-much-ct-scans...

Which isn't much for one person, but if you're CT scanning thousands of people, on average that's an extra case of cancer per thousand scanned.

The PCR test is probably more accurate and in the US, maybe even cheaper? Although it probably does have higher latency.


A lung CT can be done at about 1/5th the radiation dose of a chest CT. The future will decide if we took the correct risk with extremely limited testing in the early stages.

In my experience, “not my job” , “too much risk”, and “need more data to make a decision“ is the default attitude in the medical profession. People choose that career because it is safe, and very few doctors have military experience these days. That being said, people adjust to a new reality very quickly.


By the time pneumonia has occurred it’s too late for screening. This isn’t replacing rtPCR for people who are coughing.


Keeping everyone safe and antiseptic while you move a possibly COVID patient in and out of a CT scan will make the whole thing way more complicated and slow.


CT will not play a role in the diagnosis of Coronavirus in the US now that both the ACR and Society of Thiracic radiologists have come out not recommending it for first line diagnosis. This is a complex decision involving the availability of other tests, the potential for spread via contamination of CT, the need for CT for other things and the limited availability of scanners.

There is a group that will open source a great dataset first this sometime soon, as it might be useful in other countries.


Why is this, even though it was so prevalent in China as a key signal for testing?


I agree that it’s surprising. Medical establishment in the US is much more conservative.


2 weeks from now when just the _known_ infections number in tens of thousands the opinion of "Society of Thoracic Radiologists" will matter very little.


Nope, 2 weeks later and CT still not being used for screening covid


This is cool, but as a medical physics student, it has been repeatedly emphasized to me that CT scans are extremely dangerous and usually considered the "imaging of last resort" when no other method is appropriate. The risk of cancer from a single CT scan can vary from 0.01%-0.4%, bringing it near the risk of death from COVID-19:

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4635397/

> An estimated 1 in 270 women who underwent a coronary angiography CT at age 40 will develop cancer from that CT (1 in 600 men), compared with an estimated 1 in 8,100 women who had routine head CT at the same age (1 in 11, 080 men).

CT scans might just be the worst possible way to detect the coronavirus.


As the first comment there points out, a 96% accuracy doesn't mean much without knowing the false positive and false negative rates.

I have no doubt that this is good work in the right direction - but the headline alone doesn't make it a home run.


Do human doctors have a better rate? And at what cost/throughput?


This is the base rate fallacy.

    return “healthy”;
would have a 99.9% accuracy right now.


I think you have a good idea for an AI startup here.


For bonus points, have it return one out of every 1000 or so requests with "not hot dog."


Ahh, you need to use ROC AUC or something.


ROC AUC is somewhat meaningless when basically any amount of false positives is going to massively decrease the utility of the test.


I suppose in a case where like half the population had the virus those might factor less. And if the errors skew towards false positives, it could be a quick prescreen to get an enriched pool of candidates for manual checking. Definitely not a panacea yet though


Speaking of AI and coronavirus detection, some Chinese tech companies (Megvii, Baidu, SenseTime) have apparently built systems to detect fever with thermal imaging cameras and AI.

https://www.scmp.com/tech/policy/article/3049215/ai-firms-de...

https://www.bbc.com/news/technology-51717164

Seems like something that could be useful anywhere. Ideally, people would self-regulate (go to the doctor, isolate themselves, etc.), but in practice extra layers of protection seem smart. People may not immediately realize they're sick, they might go out anyway, etc.

This seems to, in China's case, tie in with facial recognition they already do (which has privacy implications), but there's no reason the rest of the world would have to approach it that way.


It's possible to have the virus without showing significant symptoms, isn't it? Either because you just recently caught it, or because you happen to be lucky enough to have only minimal symptoms. Doesn't that undermine temperature-based detection?


Yeah, for sure it can't be 100% effective.

I suppose what really matters is out of all the infected (or more precisely, contagious) people, what percentage of them have a fever. If it's 10%, then it's not worth doing. If it's even 50%, then maybe it could make a significant difference.

Fever is one of the three main symptoms (https://www.cdc.gov/coronavirus/2019-ncov/about/symptoms.htm...), so it seems like it has to at least be a significant percentage. But just how high a percentage, I don't know.


Thermal cameras already give you temperature output, you dont need fancy AI for that. Throw OpenCV tutorial Face Recognition code and you get the same result.


You can only do that in China.

The sheer amount of privacy-infringing data set is impossible in the US.


You don't need to write any data to storage. Just analyze stuff and make a loud beeping noise if someone has a fever.


IANAD but just keep in mind that CT scans isn't specific Covid-19, rather that it detects telltale damage done by the virus but also could be done by other respiratory viral disease. Only in places where Covid-19 is widespread, one might assume a positive test result indicates Covid-19 in the patient.


It shows an atypical pneumonia. I believe this appearance in combination with a lack of cultures for other pathogens which do this was how Covid was first suspected in China in December.


Is there a possible way to identify COVID-19 without a Chest CT scan? For those infected, a Chest CT is an added burden, and for those not infected, it would be unnecessary effects of radiation. However, the use of AI to identify COVID-19 within 20 seconds is a good way to use technology during a crisis. We offer awareness and CT Chest scans at https://www.andersondiagnostics.com/ for added safety.


A good heuristic is, don’t get your medical news from LinkedIn. Doubly-so if the word “AI” appears in the headline.


Isn't a CT scan kind of an ordeal to do?


Agreed... seems like the bottleneck is the cost and time to run the CT scan, not to have a doctor look at the results.


in china they were running partial ct scans: 5min per patient. Usually it's 30+


Sorry, modern CT scans are < 20 seconds. The really great scanners are 10 seconds.


had a couple of CT myself. Usually it's minutes of active scanning. But all the preparations/calibrations/whatever is done - you stay in machine at least half an hour+. They cut it down to 5.


Are you sure you aren’t mixing up CT and MRI? Our scheduled diagnostic CT slots are at every 15 minutes, of which probably 10 minutes are getting the patient on-off the table and positioned and 2 minutes are cleaning the room.

I do CT guided procedures every day. It’s seconds for data acquisition for any scanner made in the last 15 years. Even our ancient scanners from the early 2000s are this fast.


As a radiologist in a tertiary care centre operating CT scanners from four different manufacturers of various generations and age (0-20 years) I can assure you that there is no current CT protocol that requires minutes of active scanning time.

Even if you are using bolus tracking and a multiphase acquisition your actual scanning time will be far less than a minute.

We know in our centre the average patient time on the scanner is 10 minutes and we book patients every 15 minutes.


If a C- chest CT is taking 30+ minutes, something isn’t right. A standard workflow might have them every 30 minutes to allows extra bookings to be plugged in where needed, without ruining the rest of the day.

Or at least that’s the workflow in the places I’ve worked.


workflow times are far from what is technically possible, for a host of reasons.


The thing is that, not all COVID-19 patients will suffer viral pneumonia, and lung damages are only visible in later stages of infection, when you would not need a doctor to realize you're infected. What's the point of this?

EDIT: Hmm, so it seems like CT is more accurate than current test kits, which reportedly result in many false negatives. Yet, I found that other AI companies had already made their own models, with 90 ~ 97.6% accuracy[1].

[1]: https://translate.google.com/translate?sl=auto&tl=en&u=http%...


Spoiler: "X can do Y" is not the same as "random dude on LI says random company says X can do Y given Z without providing any relevant numbers"


accuracy don't mean shit. give me precision and recall too or you're just picking numbers.


Can it differentiate between all the various types of atypical infection? Because the findings of COVID are nonspecific viral pneumonia, which could literally be any atypical infectious/inflammatory process.

CT only has great sensitivity; it has terrible specificity.

Anyone have the paper?


Is it possible to detect coronavirus with NIRS (Near-Infrared Spectroscopy)? https://en.wikipedia.org/wiki/Near-infrared_spectroscopy

FWIU, the equipment costs and scan times are lower with NIRS than with CT or MRI? And infrared is zero rads?

(Edit) I think it was this or the TED video that had the sweet demo: "The Science of Visible Thought & Our Translucent Selves | Mary Lou Jepsen | SU Global Summit" https://youtu.be/IRCXNBzfeC4

Are these devices in production?


CT scan room and machine is going to need a lot of cleaning.


I've read that that's the real bottleneck--you need to clean the bejesus out of the machine if you don't want it becoming part of the problem.

The actual scanning itself only takes a moment. We do head ones and they're surprisingly fast once everything's ready to go.


I saw a summary of China's efforts, supposedly before they had enough testing available they were at one point running 300-400 patients a day through each CT machine in Wuhan.

That's one every 4 minutes.


Do you know if they did anything clever to keep the scanners clean?

I wonder if you could somehow isolate the patient rather than deep-clean the machine (ours has some exposed moving parts, which would make that time-consuming, but perhaps the non-research ones are better designed).


> isolate the patient

That's what I would do. Someone else that claims to be a radiology tech said you just need to wipe the machine down for SARS since it's spread by droplets and fluids.


Props to these China AI scientists!

Even more impressive is that they were able to develop it in about 6-7 weeks after the initial containment effort.

This is a win for AI, as a massive pattern recognition machine, to help reduce the labor load of critical personnel, like radiologists. It helps by removing them as the limiting factor, that would slow down a critical process.


A radiologist can do it in about the same time. It presents as "ground glass opacities" which are pretty easy to spot.


Is this actually detecting COVID-19 or yellow rulers?

https://www.ben-evans.com/benedictevans/2019/4/15/notes-on-a...


Would be awesome if radiology finally gets disrupted. 50% of hospital bills can be attributed to the radiology department thus scalable innovation will result in a more affordable healthcare service all round.


It would be good but not sure that radiologists are the dominant cost. Radiological equipment is heavily regulated because it can cause significant harm to people if the equipment malfunctions.


I think this is the case when false positives and false negatives can do a lot of harm, so in human review of CT scan is a must. Especially if we consider how small dataset this model is trained on.


Only if you are way past symptomatic and have bilateral pneumonia. The CT scan wait time is extra.


Github, or it didn't happen




Consider applying for YC's Spring batch! Applications are open till Feb 11.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: