Thanks to Brilliant for supporting this episode of SciShow.

Go to Brilliant.org/SciShow if you're interested in investing in your STEM skills this year. [ ♪INTRO ]. If you've ever ended up with a nasty rash from using skincare products, especially oily or heavily scented ones, you're not the only one.

A lot of people react to certain compounds found in these products. Around 50% of people who use these products will experience this allergic reaction, known as contact dermatitis. But researchers may have finally figured out why these pesky rashes happen — and how to prevent them.

Right now, the main way to treat contact dermatitis it is to just avoid products containing certain chemicals. But if you've ever had this problem, you know that's a long list. And before now, scientists simply didn't understand how these rashes happen.

See, allergic reactions are often triggered by specific molecules called peptides. Those peptides trigger immune cells known as T cells. But skincare products don't typically have those kinds of peptides in them.

What's more, the molecules they do have are thought to be too small to be seen by T cells. But last week in a paper published in the journal Science Immunology, researchers showed that a molecule found in our skin called CD1a binds to certain skincare chemicals, making them visible to T cells. It basically rats them out to our immune system.

The researchers identified several common skincare substances that were able to cause a T cell response by binding with CD1a. Two molecules found in a commonly used vanilla-scented oil, benzyl benzoate and benzyl cinnamate, got T-cells fired up when they were bound with CD1a. The researchers looked even more closely at another allergen known as farnesol.

They found that rather than just sitting on the surface of CD1a, it actually binds deep inside it , displacing natural skin oils that would normally be there. That meant T cells weren't simply recognizing the chemical structure of farnesol on CD1a, but instead changes to the shape of CD1a itself. The researchers believe they might be able to identify other compounds that can compete with farnesol for a spot binding to CD1a without causing an immune response, offering some hope for preventing contact dermatitis.

Another idea getting a lot of attention this month comes from a pair of papers that claim to show how artificial intelligence can be trained to detect cancer more efficiently than doctors. The first study, published last week in the journal Nature, outlined an algorithm for detecting breast cancer from mammograms, which are essentially x-rays of breast tissue. Researchers first trained the AI to recognize cancer by showing it tens of thousands of mammograms from women in the US and UK with a confirmed diagnosis.

They then tested the AI on different datasets of around 26,000 UK women and 3,000 US women and compared its results with the initial diagnosis made by expert radiologists. The algorithm caught cancer on images where it had been missed by the doctors who initially examined those mammograms. That will be a false negative,when we are saying it isn't there, but actually is.

And it reduced false negatives by 9.4% for the US dataset and 2.7% for the UK dataset. UK doctors always get a second opinion, which might help explain the difference. Which is great, because doctors can miss up to one in every five cases of breast cancer.

So even a few percentage points could be helpful. The AI also lowered the rate of false positives — where it looks like cancer is there but it actually isn't — by around five percent for the US group and one percent for the UK group. The second study, published this week in Nature Medicine, used a similar method to train an.

AI to detect brain cancer. Researchers trained their AI on a dataset of 2.5 million images of brain tissue from several hundred patients. But when it came to testing the AI, these researchers did it in real time.

They took two samples of brain tissue from 278 patients during surgery and gave one to their AI and one to a team of pathologists. The computer would first create detailed images of the brain tissue and then analyze them using an algorithm. The humans would go off to the lab and look at the samples the old-fashioned way, using a microscope.

The AI did slightly better than the experts here too, getting the diagnosis right 94.6% of the time compared to 93.9% of the time for the humans. But what was really amazing about this technique was its speed. The AI could predict brain cancer right there in the operating room in around two and a half minutes, instead of roughly 30 minutes it would take humans to do it.

Which is great, because when doctors are operating on your brain, they want to be really sure about their diagnosis. Now these studies don't mean cancer diagnosis is solved forever. One concern about algorithms in general is that they're only as good as the dataset they're trained on.

So if the dataset doesn't include people of different races or sexes, then the AI might not work as well for those groups of people — like if the disease manifests itself differently in, say, men with breast cancer. They also wouldn't work for diagnosing rare tumors because there isn't enough data to use for training the algorithms. Plus, it's currently unclear exactly how to use AIs in real-life hospital scenarios.

Doctors could use them alongside their own expertise to help them diagnose disease. However, a 2015 study suggested that other computer-aided methods of detecting breast cancer didn't improve accuracy, and may even have made things worse. In addition, some physicians have raised questions about these studies, suggesting that we should be identifying patients with dangerous but curable cancers, not teaching AIs to find as many as possible.

Especially when those lesions could be harmless, leading to unnecessary treatment. In other words, the question no longer seems to be whether we can train AIs to help us find cancer, but how they should be used to do that. If you're interested in understanding more about how the AIs we talked about today can detect cancer, you might like the computer science courses over on Brilliant.org.

They even have a machine learning course, which will show you how computers learn, and why training them is so important. In addition to computer science, Brilliant offers courses related to science, engineering, and math. They're a great way to understand the world better in the new year.

They're designed by talented educators and lifelong learners from top institutions, so you know they know how to make things stick. And courses are even available offline via Brilliant's iOS and Android apps. The first 200 people to sign up at Brilliant.org/SciShow will get 20% off an annual Premium subscription.

And by checking them out, you're also supporting SciShow — so thanks! [ ♪OUTRO ].