AI Built to Identify Pastries Recognizes Cancer with 99% Accuracy

  • Who would’ve thought a croissant could end up helping us combat cancer?

Sometimes, the best inventions are accidental. A Japanese AI that originally had nothing to do with medicine could just inadvertently made a great contribution to global health.

BakeryScan is – like its name implies – a computer AI system for bakeries. Developed in 2013 by computer engineer Hisashi Kambe, bakeries use the AI to identify different pastries from each other.

The system makes the bakery checkout process more hygienic and helps cut down on staff training costs. It’s a fairly common solution that any bakery can purchase for $200,000.

BakeryScan’s AI learns as it keeps identifying more and more pastries. But no one expected it to learn how to identify cancer.

This isn’t a one-off accident, either. BakeryScan is good at spotting the insidious disease – really, really good.

In fact, it can identify cancer cells at a 99% accuracy. That’s about as good of a result that you’re ever going to get.

“I’ve analyzed 1,264,963,136 pieces of bread. I can tell this one has cancer.”

Saw It on the TV

But how did a bread-identifying AI suddenly get a medical degree? Turns out, the right TV ad in the right place can make all the difference.

BakeryScan’s transformation from a bakery tool to a medical aid began in 2017. That’s when a doctor from the Louis Pasteur Center for Medical Research in Kyoto saw a TV segment about the BakeryScan.

Something unusual in the ad caught his eye. The highlighted bread products in the ad looked an awful lot like what cancer cells looks like under a microscope.

This revelation gave the doctor an idea. Could this machine be modified to identify cancer instead of bread?

He contacted BRAIN, the company that makes BakeryScan. After a bit of negotiation, BRAIN agreed to start developing a medical version of their system.

The whole thing worked out pretty well. BRAIN had already developed a way for BakeryScan to find “interesting features” in images, and it was able to learn from human feedback.

Soon afterward, Cyto-AiSCAN was born. It’s a medically optimized AI system using the same principles that tell pastries apart.

Currently, the AI is being tested in two major hospitals in Kobe and Kyoto. And it’s doing a bang-up job.

A Serving Tray of Cells

It’s no wonder, though, that BakeryScan is so good at spotting cancer cells. The process to identify bread and cancer is surprisingly similar.

In a bakery, BRAIN’s AI-SCAN technology casts its electric eye over a tray of pastries. It recognizes each unique pastry type and gives it the correct price.

As a result, bakery staff can quickly move the breads and pastries out for sale. The whole process is much faster than the staff guessing what each kind of pastry is and then trying to remember the right price tag.

Like many other AIs, BakeryScan learns as it works. If it misidentifies a pastry, a staff member can tell it that “no, that’s actually wrong,” which will leave an imprint in the AIs memory.

Over time, BakeryScan has gotten extremely effective. It can now even tell whether a pastry has been damaged, and it no longer mistakes two touching pastries as one item.

At a hospital, the whole thing works pretty much the same way. Only, instead of a tray of pastries, the AI now scans a microscope slide of cells.

But just like anybody, the AI had to learn how to do its new job from scratch. Initially, it was only able to examine one cell at a time and the results were sometimes a bit so-so.

AI-SCAN is a quick learner, though. With input from human experts, it began to pick up the speed and accuracy of the scan.

Now, it can analyze an entire slide of cells with one glance instead of having to zoom in on each cell for a closer look. And, like we said, it’s fast approaching 100% accuracy.

Lifelong Learning

Inspired by the results of the Cyto-AiSCAN, BRAIN has now started developing its software in other directions as well. It seems whatever the system needs to count, it’s great at it.

AI-SCAN’s algorithms have been used, among other application, to count and identify pills in hospitals. Outside the medical field, it has counted the number of people in classic Japanese woodprint block and sorted charms and amulets at shrines.

Even in bakeries, the BakeryScan keeps on learning. When the COVID-19 pandemic hit Japan, bakeries started wrapping their products in cellophane to keep them sterile.

Unfortunately for BakeryScan, the reflections of the plastic wrapping threw off its algorithm. But a solution soon came up – bakeries started showing the AI stock pictures of their products in a plastic wrap.

Soon enough, it learned to ignore the plastic surface and look a bit deeper. And there they were, the same familiar pastries the AI had gotten so good at recognizing.