Slashdot Mirror
Deep Learning Algorithm Diagnoses Skin Cancer As Well As Seasoned Dermatologists (extremetech.com)
An anonymous reader quotes a report from ExtremeTech: Remember how that Google neural net learned to tell the difference between dogs and cats? It's helping catch skin cancer now, thanks to some scientists at Stanford who trained it up and then loosed it on a huge set of high-quality diagnostic images. During recent tests, the algorithm performed just as well as almost two dozen veteran dermatologists in deciding whether a lesion needed further medical attention. The algorithm is called a deep convolutional neural net. It started out in development as Google Brain, using their prodigious computing capacity to power the algorithm's decision-making capabilities. When the Stanford collaboration began, the neural net was already able to identify 1.28 million images of things from about a thousand different categories. But the researchers needed it to know a malignant carcinoma from a benign seborrheic keratosis. Dermatologists often use an instrument called a dermoscope to closely examine a patient's skin. This provides a roughly consistent level of magnification and a pretty uniform perspective in images taken by medical professionals. Many of the images the researchers gathered from the Internet weren't taken in such a controlled setting, so they varied in terms of angle, zoom, and lighting. But in the end, the researchers amassed about 130,000 images of skin lesions representing over 2,000 different diseases. They used that dataset to create a library of images, which they fed to the algorithm as raw pixels, each pixel labeled with additional data about the disease depicted. Then they asked the algorithm to suss out the patterns: to find the rules that define the appearance of the disease as it spreads through tissue. The researchers tested the algorithm's performance against the diagnoses of 21 dermatologists from the Stanford medical school, on three critical diagnostic tasks: keratinocyte carcinoma classification, melanoma classification, and melanoma classification when viewed using dermoscopy. In their final tests, the team used only high-quality, biopsy-confirmed images of malignant melanomas and malignant carcinomas. When presented with the same image of a lesion and asked whether they would "proceed with biopsy or treatment, or reassure the patient," the algorithm scored 91% as well as the doctors, in terms of sensitivity (catching all the cancerous lesions) and sensitivity (not getting false positives).
Wow, both cancer AND seasoned dermatologists!
I'm sure it can diagnose the heavily seasoned dermatologists, but how does it handle the merely lightly seasoned?
...but they aren't programmed by evolution to disregard 95%+ of it. Pretty much the exact opposite actually. We tune, prune, select, and evolve these algorithms to do this one thing really well. Frankly, its a wonder humans can do as well as we do. A testament to our pattern matching skills, adaptability, and lack of immutable hard wiring in the 'ol thinky thinky bits.
Something I saw that might be able to help humans get a step up on the algorithms, or actually amalgamate humans and computers, is this:
Video Magnification
Maybe cancerous skin lesions absorb slightly different light wavelengths? If so, magnification of the minuscule differences could pinpoint it. Fun to ponder.
When the only tool you have is a claw hammer every problem starts to look like the back of someone's skull.
I have had both Basal and Squamous skin cancers since the 1990's and keep a close watch on my own skin. If I see anything suspicious I have a note book where I keep a note of what I saw, when and where. In some cases I will take a close up picture of it. Both Basal and Squamous cancers tend to appear and go away when they are very small and by doing this I have a record of "something" reappearing in the same location. Following the old adage that once is happenstance, twice is coincidence, but three times is most likely enemy action I will call for an appointment with my dermatologist and show them my records or pictures. For the last ten years I found every skin cancer well before the dermatologist would have seen it during an annual exam.
It did not used to be that way since for many years I had the same dermatologist or group and they got to know my skin about as well as I learned to. However, after that with almost yearly shifts in medical networks due to changes in insurance providers where I worked (always either the lowest bidder or highest campaign contributor), it got where I didn't see the same one twice until I got on Medicare. The patient-doctor relationship SHOULD be long term and more than just a diagnostic code and EMR's. I think it is going to get a LOT worse before it gets better so learn to know your own body and be assertive about your care.
It's all insurance providers. Hopefully AI can detect them as a cancer on the system and take appropriate action, because human politicians with their eyes on campaign contributions sure as hell can't.
"Transparent" is a shit show that trades on every stereotype going. A man in drag is NOT a transsexual.
Reading an article like this, I can just hear my regulatory affairs officer having a heart attack in my head.
I realize that very few people here have ever had dealings with the FDA. The FDA regulates the interstate marketing of medical tools (including software) and drugs, everything they do comes from that core mission and authority. Press releases and statements are pretty central to that mission. You should try to limit press on your product to what the FDA agrees you've proven. Depending on your views of the government, medical ethics, and your risk tolerance, "should try to" in that statement might be "must" or "should pretend to."
Software focusing on "health" isn't really regulated, but "diagnostic" means this is medical. If you think they made a diagnostic tool for skin cancer, then they may have a problem when it comes time to talk with the FDA. They haven't shown that they have a diagnostic yet, that's the point of the last quote in the article, but that quote is in FDA-speak while the rest of the article is less formal sounding. (They've done what's called a retrospective study, which is at most half of what is necessary.)
Generally, the authors on the paper are smarter than this. Here's an example of an article a Stanford Dermatologist usually contributes to. Note that the Dermatologist quoted in that article is also quoted in TFA. Note the difference in tone of the publication, the whole thing is in FDA-speak. Yes, it's super boring. It's also not going to give anyone at the FDA a reason to hold up an application for marketing prior to approval.