From the bakery to the laboratory: The future of AI in cancer recognition

BakeryScan, an artificial intelligence (AI) designed to recognize different types of pastries, has made its way into the labs of cancer research. The same technology used to differentiate between bagels and croissants is now being tested to identify cancer cells [1].

Computer Vision, or Lack Thereof

The ability to see is one we often take for granted. Humans can look at a bouquet and identify what flowers are in it. We can look at a piece of fabric and determine its texture, or a wall and estimate how far we are from it. Computers, on the other hand, are limited to pixels and their respective colors, a series of codes in place of light. Identifying shapes does not come as naturally nor as quickly to a computer as it does to a human. As such, image recognition became one of the biggest challenges in AI.

In the early 20th-century, the concept of AI was already formally proposed by a few [2,3]. However, it would take some time before computers would even be capable of their “own thought.” This presented one of the first problems with regard to image recognition: early computers were incapable of carrying out the necessary processes. Moreover, it was extremely costly. Simply running a computer used to cost upwards of a hundred thousand dollars [4]. It would take decades of work before computers would be able to store more information and work more efficiently.

Today, AI has gone a long way from being part of fiction books. They have become a daily occurrence. There have even been instances where an AI was able to beat a human at games such as Chess and Go [5,6]. More importantly, they were now capable of carrying out image recognition. Thus, a second issue arose: how would computers analyze the information they were given?

Various methods were attempted such as the block world [7] and polyhedral modeling [8], but they were never as accurate nor as detailed as engineers wanted. Then came deep learning, which works on a computer as a brain does to a human–it makes millions of connections at a time with computing cells, akin to neurons in a brain. When applied to image recognition, deep learning takes the vast web of images and compiles them to create a database to reference new images against the collection [1,9]. For example, given a picture of a cat, the computer would reference its database of other cat images in order to determine that the picture is, in fact, a cat.

This method, however, had a limitation: it requires thousands of images to comprise a database which takes quite some time to compile. Thus another approach was made, this time starting with a bakery in Japan.

The Birth of BakeryScan

Pastries are a big deal in Japan. You are as likely to find pastries as you are to find rice in a Japanese household [10]. As such, bakeries take great efforts to profit from this. In a market study, it was found that more products were sold if they were shown on display and without packaging [1]. With no packaging and subsequently no barcodes, cashiers were forced to memorize every variety of pastry and their price. This system was inefficient and unsanitary, so one bakery decided to approach BRAIN, a programming company headed by Hisashi Kambe, to come up with a solution.

There were several things to consider — the difference in lighting, the presence of crumbs, the holes in the bread. Each and every aspect had to be accounted for in order to accurately identify a pastry. The need for highly accurate and controlled conditions, alongside the 2008 financial crisis, caused BRAIN to struggle. After years of trying, in 2013, BRAIN came up with BakeryScan, a system that was able to accurately differentiate between 50 different types of pastries 98% of the time. This works by going through a series of checkpoints [1]. For instance, one check would contain an algorithm to adjust coloring, while another would analyze if pastries were placed too close to each other, and so on. This was groundbreaking because, unlike deep learning, BakeryScan only required around twenty images to start accurately identifying the item — a total increase in productivity.

Now, BakeryScan is worth about 20,000 US dollars and is deployed across hundreds of bakeries across Japan. With the amount of potential it holds, someone was bound to take notice.

What’s Next?

But are pastries the only thing BakeryScan is limited to? The AI’s capabilities were recognized by a doctor at the Louis Pasteur Center for Medical Research in Kyoto back in 2017. The doctor noted that cancer cells looked remarkably similar to bread. It is here that the journey for BakeryScan in the field of pathology began. Currently, the system could detect cancerous cells based on the cell’s size, texture, roundness, nucleus, etc., with a 99% accuracy rate. Additionally, the system adopted a new name: Cyto-AiSCAN.

Today, the program continues to be tested in two major Japanese hospitals for its cancer-cell detection capabilities. The company also claims that the AI may be used to distinguish pills in hospitals, detect poorly wired bolts in jet engines, and count figures in paintings [1,11].

With its humble beginnings in a bakery and its future in cancer research, the AI’s story truly is an inspiration for all. Cyto-AiSCAN is indeed one to keep an eye on, but in the meantime, let us appreciate where it all began — the BakeryScan.

References

[1] Somers, J. (2021, March 18). The pastry that learned to fight cancer. The New Yorker. https://www.newyorker.com/tech/annals-of-technology/the-pastry-ai-that-learned-to-fight-cancer

[2] Turing, A. M. (1950). Computing machinery and intelligence. Mind, 49. 433–460. https://www.csee.umbc.edu/courses/471/papers/turing.pdf

[3] Gugerty, L. (2006). Newell and Simon’s logic theorist: Historical background and impact on cognitive modeling. Proceedings of the Human Factors and Ergonomics Society Annual Meeting. https://doi.org/10.1177/154193120605000904

[4] Anyoha, R. (2017). The history of artificial intelligence. SITN. https://sitn.hms.harvard.edu/flash/2017/history-artificial-intelligence/

[5]Sneed, A. (2016). Computer beats go champion for first time. Scientific American. https://www.scientificamerican.com/article/computer-beats-go-champion-for-first-time/

[6] History.com Editors. (2021). Deep blue defeats Garry Kasparov in chess match. History. https://www.history.com/this-day-in-history/deep-blue-defeats-garry-kasparov-in-chess-match

[7] Szeliski, R. (2010). Computer vision: Algorithms and applications. Academia. https://d1wqtxts1xzle7.cloudfront.net/54343495/Algorithms_and_ApplicationsSzeliskiBook_20100805_draft-with-cover-page-v2.pdf?Expires=1635860229&Signature=YlcNu6Q8CmFDnQQsCjTUBhwD1eRKKvQMvLkpz8Hp0hYxRl~172yKTRXafthIU27~p3Z6MNDTT4FPPwTs5cxme0JeJfE~5v5y-sRm1-OARdfxeW9fQVpx4pFoC5SnRXn5X3zioGr1PnqdfHwbqxRf4sryuljPHd1cL0ZZxNglFAVQbeHXmmvxRg4Or0HU1p7jqBVl~60xEAkisCIjhPZqtsEZbqO5pyDKXb589FVw7QlVbpxIdYS5hGVx2VVWcYrF~GVhG84-880iZk1LTuuWXdWgxlIx6SOLjlEnGMwW4yaHfA2Ta1kE0RyVeS6p-aAnBBmO3taR1um8Ha4-sU67tg__&Key-Pair-Id=APKAJLOHF5GGSLRBV4ZA

[8] Baumgart, B. G. (1975). A polyhedron representation for computer vision. National Computer Conference. 589–596. https://people.cs.clemson.edu/~dhouse/courses/405/papers/p589-baumgart.pdf

[9] O’Mahony, N., Campbell, S., Carvalho, A., Harapanahalli, S., Hernandez, G. V., Krpalkova,L., …, Walsh, J. (2019). Deep learning vs. traditional computer vision. Advances in Computer Vision, 1. 128–144. https://arxiv.org/pdf/1910.13796.pdf

[10] Rana, P. (2020). Rising success with bakery goods in Japan. Tokyoesque. https://tokyoesque.com/bakery-goods-in-japan/

[11] BRAIN Co. Ltd. (2016). Products. http://corp.bb-brain.co.jp/english/

Written by Regina Clare Ruby
Proofread by Sophia Abulencia, Megan Gozum, Gabbie Lagdameo, Via Ogasawara, and Rhaena Pablo
Art by Regina Clare Ruby