Hacker News new | comments | show | ask | jobs | submit login
Low-cost, non-invasive melanoma detector wins award (jamesdysonaward.org)
312 points by ductionist 4 months ago | hide | past | web | favorite | 50 comments

The creators are from McMaster* University in Canada, their names are Rotimi Fadiya, Prateek Mathur, Michael Takla and Shivad Bhavsar. Really wish they got the credit in the article.

questions I have from a clinical perspective: Today, we see skin lesions that look like melanoma and apply the ABCDE acronym for diagnosis:


Border irregularity

Color (all the same or not)

Diameter greater than 6 mm

Evolving size, shape or color (over time)

Now, based on this, we would then probably biopsy the lesion (any dermatologists lurking on HN wish to chime in?) and send to pathology to interpret.

Will this method beat out the clinical diagnosis in terms of sensitivity (number of false negatives, which is what I would care more about) and specificity (number of false positives), such that we could biopsy more of the patients that we missed before? Additionally, would we see these changes earlier with this software and hardware?

Cool idea, nonetheless. I'd be more than happy to set up some sort of clinical study looking at this technique though if there are enough tools out there to implement this myself.

Not a dermatologist but I've following melanoma research for several years now thanks to my wife's diagnosis.

What's exciting about this one to me is that if the detection mechanism is truly sound, this is a great leap over the ABCD(E) criteria which have been the clinician's main tool, and which has a high false negative for one of the deadliest forms of melanoma, nodular melanoma. This often does not present with varigated coloration and because it tends to invade tissue rather that spread superficially its morphological characteristics aren't as apparent until its more developed, greatly increasing the risk of metastases. This was the case with my wife's primary, which her dermatologist almost didn't send out for a pathological workup, and had been missed by a couple of different doctors thanks to its atypical presentation.

All of which is not to say that ABCDE isn't a valuable rule of thumb, but its always struck me as somewhat subjective, which I guess is tempered with the experience of the physician but also does have the false negative problem, even with specialists.

Best of luck to you and your wife.

ABCDE is awful, don't get me wrong. I've always felt it was like holding your finger out to see what the wind direction is like. It has a horrible sensitivity (50%-90% depending on the type, I'd err on the side of lower). But if this technique also only has a 50% sensitivity, that is what would bother me. Hopefully the sensors they are using are sensitive enough to detect these temperature changes well.

Thank you, her prognosis is very good—we caught it just early enough. There was a micrometastasis to the sentinel node but no further spread, and we're coming up on the 5 year mark next year.

Agree wrt sensitivity—obviously while this is novel and exciting to me at least, it bears further study before becoming part of the standard of care.

> [ABCDE] has a horrible sensitivity (50%-90% depending on the type). But if this technique also only has a 50% sensitivity...

If the sensitivity of the two techniques are not correlated, a 50% improvement on top of ABCDE would be fantastic.

Then call it ABCDEF (Frigid cells warm quicker)

Thanks for your insight. Do you know how well the Molescan system compares with manual identification? My wife has had a few suspicious moles flagged and removed (none tested positive yet, thank goodness).

I don't know much about the efficacy of those systems, I just looked into patents on diagnostic systems for melanoma awhile back—there's a dermatologist elsewhere in the comments who might be able to speak with some authority about it. My sense is that such systems are primarily for surveillance, detecting changes over time. I suspect that in a controlled study against standard photographs of your body over time (the kind you can do at home) both methods are equally useful. The variable would be how diligent you are in taking those photographs—if you do those with the frequency one should, and cross referencing photos over time, my guess is you're as likely to catch something as what you'd get out of a regular series of visits to a Molescan clinic.

Hi Dermatologist here

These features (ABCD) are very helpful for patients and other doctors in highlighting moles that should be checked. However, dermatologists generally use a dermatoscope (magnified image) to examine the pattern of pigment within the mole as an important part of deciding whether it is benign or malignant.

There is a long list of dermoscopic criteria that can help to identify melanoma and this is used by doctors training in dermatology but in practice I think that it is seeing lots of normal moles that is the most useful. The brain becomes trained to recognise a normal mole and when it sees something that does not match this pattern it is labeled as suspicious - often I cannot say exactly why.

Recognizing a clear cut melanoma is very straightforward, even for a medical student the difficulty is in distinguishing moles that look abnormal but are actually harmless. At the present time these would be excised and examined under the microscope and as a consequence a large fraction (up to 90%) of moles removed are harmless moles.

I suspect that this device will not increase the number of melanomas diagnosed, but if it is sufficiently sensitive and specific it might reduce the number of normal moles removed.

I have had two melanomas and might have my third (just got some preliminary results back).

I have more than a thousand moles and is lucky to be at Sloan Memorial. They found my second one by pointing some laser onto the mole without biopsy and could see the malanoma there. I have 3d images of me and all sorts of things.

This sounds more like something I would need (given my high number of moles) if I am lucky enough that this third one either isn't a melanoma or early stage.

One thing to note out which I learned. Asymmetry needs to be understood a little different. I have a lot of asymmetric moles that are just fine. They can even grow and it will be fine as long as their internal network nodes grows the same (think of it as a vector image that will grow uniformly.

What's the name of the laser procedure you had?

Can't remember but I can find out. It had to be strapped on to me very precisely and they had a lot of issues with keeping it steady. Once it was one we could actually see the cells.

likely confocal microscopy

I just read today that in vivo confocal microscopy was used clinically in human patients. Do dermatologists do this, or is someone else operating the device? I've only used them in mice, very neat!

Yes in Vivo confocal microscopy is done by Dermatologists

The machine is called vivascope

There's a great New Yorker article [1] from April that talks about this specific example for trying to make machine learning methods to spot melanoma from pictures. What I learned (as a non-physician) is that ABCDE is what you tell the students but after seeing hundreds of examples in person, the physician does it by some impossible to describe criteria that are far more nuanced. The question really is can a machine out perform that impossible-to-describe function? You might find it interesting.

[1] https://www.newyorker.com/magazine/2017/04/03/ai-versus-md

I've read the article :), the author is the same physician who wrote "emperor of all maladies", a book about cancer.

Basically what it comes down to is that doctors get really good at bayesian inference at some point in their career, definitely not while we are students. I am however confident that we can build models that learn in a similar manner.

I'm actually trying to build a model for a class project that uses bayesian inference in a CNN as per Gal et al (https://arxiv.org/abs/1506.02158) to try and detect diabetic retinopathy. Their paper actually tries to detect different skin cancers.

Miiskin tries to develop such a recognition algorithm [0]. To get the training sets they offer a app which allows you to take pictures, periodically reminds you and keep track of moles. The pictures are uploaded outside EU and linked with PII like email, sex & year of birth. Including possible pictures of your face, breasts and genitals [1].

I don't want to pay with my personal data. Is such an app recommended and are there open source alternatives?

[0] with the purpose of enabling our Image Pattern Processing system in 3.12 http://miiskin.com/legals/

[1] https://youtu.be/pTzk7rfcc-I

I'm not sure that the app is "recommended" per say, but I guess it wouldn't hurt to be extra cautious if you're out in the sun a lot or predisposed.

The only other app I know is SkinIO, developed by some northwestern students if I recall correctly. I'm not certain how it all works, but maybe look into that?

Sounds like a really good match for ML - as long as all the input the physician is detecting is visuals. There could be e.g. a smell or temperature component as well, but it's likely visual is enough to get very high accuracy eventually.

Will it beat clinical diagnosis in the first world, maybe not. But if it's cheap, it could be made available gloabally and potentially save a lot of lives.

Are there affordable treatment options for folks with skin cancers in non-1st wouuld countries? If not, then this test seems like it is essentially a "congrats, you have cancer, now you and your family will know what kills you"

Melanoma when caught early is very easily treated—you simply cut it out, which has a 98% cure rate. Once it starts to invade subdermal tissue or spreads to the lymphatic system, outcomes rapidly become less optimistic. As recently as a few years ago, a stage III diagnosis had around 50-60% 5 year mortality, and a stage IV had a 90% 2 year mortality. Because melanoma is primarily an immune-system mediated cancer, many of the recent advances with targeted genetics and immune-system stimulation help melanoma patients directly, and the "new" numbers on those cases aren't able to be calculated yet simply because enough time hasn't passed. But the general consensus is that they're going to have to revise those outcome predictions a great deal now that we have these tools available.

The question of how many of those treatments are affordable is largely driven by the country they're in—many countries regulate the cost of pharma in general, so they'd be more affordable than in the US.

I'm not an expert and this may be simplifying, but I did know a guy who was diagnosed with melanoma. The treatment was to cut a patch of skin off and wait for it to grow back.

If someone from Watsi sees this, I'd love to know if this is something they'd work with their clinical partners to deploy.

Can the initial check (diagnostic) be done with camera app and update the capture image(s) to the cloud and come back with "No problem" or "Check with your doctor soon" kind of message?

Anyone know what kind of potential regulatory/FDA issue will this kind of apps have?

I think you would just have HIPAA laws for protecting patient data (in the US) to deal with. Since you aren't making an actual decision on someone's care, I don't think the FDA would really care.

I've seen multiple attempts to get melanoma detectors past the FDA. The last attempt cost millions of dollars on appeals and still ultimately failed.

Hopefully thia one will do better.

It seems to me that the FDA should allow the best available option through even if it's not very effective, as long as it's safe and no false claims are made. This would at least allow companies to bootstrap and iterate if their approach works at all.

I believe FDA criteria is it has to be as good or better than the best available technique, which AFAIK is a detailed exam by a trained dermatologist, possibly using several tools like light sources, magnification, etc.

Could you do this with looking at post-cooling images from something like FLIR One? http://www.flir.com/flirone/


If I understand correctly, the process they're using is something like the following...

1) Cool the skin down to some depth by applying an icepack

2) Remove icepack and take a video of the skin with infra-red (a temperature map) as the skin reaches thermal equilibrium.

3) Process the "stack" of images by measuring the thermal conductivity at each point-- you do this by getting a curve of the temperature for each point.

This will give you a map of structures underneath the skin which have different material properties (a different thermal conductivity).

Perhaps, but from the video it appears they have some software that helps with diagnoses as well. Something the Flir camera wouldn't have (at least not specific to skin cancer).

Not trying to belittle the project, it's very cool - just might make sense to use existing commodity hardware. Yes the software would be crucial - FLIR One + Android/iOS software for analysis would be a quick route to market

I think a proper blood test for proteins should detect cancer e.g. https://en.wikipedia.org/wiki/Alpha-fetoprotein test to detect liver cancer

Did they invent some new mechanism to detect melanomas? Or did they just automate existing methods?

This is a novel method—existing patents I've seen on melanoma scanners generally fall into the category of 3D imaging and computer vision to automate what a clinician does to various degrees.

So is it basically that the specific heat capacity of the material in the top layers of the skin is altered when there's melanoma present, so when it's cooled one can observe that this part warms more slowly (?) than the surrounding tissue (because of density?).

It seems likely to me that there would be a difference in reflectivity/transmittivity too that could be detected by using a strong light source that could penetrate the exterior dermal layers. Presumably active scanning like MRI (or microwave heating - probably too risky) would show the differences well too?

Would be interesting to read about the range of methods that have been tested (conductivity using surface contact probes would seem another obvious method?).

The benefit, as it appears, of the current method is probably simplicity but it doesn't appear to scale to scanning the full surface readily.

The article says that cancer cells have a “higher metabolic rate” which I understood to mean that they use more energy and thus generate more heat.

My wife works at McMaster. Lots of great research happening there.

Can you give some examples?

Regardless of the usefulness of the device $40K award doesn't seem a lot, especially for a company like Dayson.

I wonder why the down-votes? For a company with hundreds of millions if not billions stashed in tax heavens and a founder who's worth £2.5bn to give a $40k award a year seems like an insult to me. It's not bad but still...even McDonalds had better campaigns as far as social responsibility is concerned.

Let's not forget he was a proponent of Brexit

I read this headline as "Mellodrama Detector" X-D


Wait, they’re a scam, but you think there should be more of them?

GP is a little jaded IMHO (but perhaps with some reason, if family members have dealt with melanoma—my experience has been different) but essentially spot on WRT to the supply and demand. The medical profession is protectionist to the extreme. Basic economic forces should see the supply of medical schools (and doctors that graduate from them) increasing, but the AMA keeps a tight lid on that, and specialization boards further distort the market.

Probably meant to say that the licensing is a scam.

Applications are open for YC Summer 2018

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