Old Google algorithm, taught new tricks to help diagnose lung cancer

AI is increasingly used across the tech industry for multiple tasks, from speech and face recognition to spam filtering or, in the medical industry, performing doctors’ work in only a fraction of the time humans need.

AI usually refers to machine learning or neural networks modeled after the human brain.

These neural networks can be trained using real data to make them able to spot a spam email, transform your spoken words into text or recognize shapes, people, locations etc.

In the September 17th issue of Nature, scientists from New York University published their new research about a Google deep learning algorithm that was trained to tell the difference between two of the most common types of lung cancers.

Repurposing AI to identify genetic mutations in cancer cells

Repurposing AI, similar to what other companies are doing with drugs, seems to be working. The type of AI now used to identify faces, animals, and objects in pictures – which are then uploaded to Google’s online services – has been used for diagnosing other conditions, including diabetic blindness.

One new version of AI, NYU’s neural network, is doing something nobody expected to be possible.  Even the scientists involved don’t understand how it works. The computer networks learned to do something that even the most experienced pathologist can’t do without further testing: from a picture, the neural network is able to identify the genetic mutations happening inside each tumor.

Since, according to the American Cancer Society and the Cancer Statistics Center, more than 200,000 people are diagnosed with lung cancer each year, and more than 150,000 people die annually as a result of disease-related complications, these findings bring really good news.

Why results are so promising for future diagnosis

Adenocarcinoma (LUAD) and squamous cell carcinoma (LUSC), the two types of lung cancer AI was able to differentiate, are considered difficult to distinguish without confirmatory tests. Usually, highly experienced pathologists are required to examine the samples to correctly determine the type of cell. The machine learning program managed to reach a 97 percent accuracy in differentiating the two types of cancer, when “reading” pictures of tumors from The Cancer Genome Atlas.

Scientists then trained the network to predict mutated genes in LUAD. It was proven that six of them, STK11, EGFR, FAT1, SETBP1, KRAS and TP53, can be predicted from pathology images. Depending on the gene, the accuracy of the determination varied between 73 and 86 percent.

Targeted treatments now come in various shapes and forms. The biotech industry is very interested in developing more targeted therapies to help patients suffering from various types of cancer and other diseases. Since the identification of altered genes is considered significant for the development of targeted treatments destined to only act against cancer cells with specific mutations, this is very encouraging news.

Response time is another point in favor of the new AI tool. While existing tests to detect such mutations need weeks to deliver results, the new AI tool reportedly provides them instantly.

When it comes to cancer, sooner is never too soon to receive the right kind of therapy.

“Delaying the start of cancer treatment is never good,” says senior study author Aristotelis Tsirigos, PhD, associate professor in the Department of Pathology at NYU School of Medicine and NYU Langone Health’s Perlmutter Cancer Center. “Our study provides strong evidence that an AI approach will be able to instantly determine cancer subtype and mutational profile to get patients started on targeted therapies sooner.”

This is not the first time google delivers important medical tools

Google has an entire history, recent, but also more distant, of providing tools for various purposes, medical or security related.

Google subsidiary DeepMind has demonstrated a machine learning algorithm that can recommend treatment for more than 50 eye diseases with 94 percent accuracy.

On September 3rd, Google released a free artificial intelligence tool that also analyses images, but for a different purpose. It will help companies and organizations identify images of child sexual abuse on the internet.

Google’s Content Safety API uses deep neural networks to process images in such a way that fewer people need to be exposed to them. Using this technique will increase the percentage of child abuse content spotted by 700 percent, Google said.

“Quick identification of new images means that children who are being sexually abused today are much more likely to be identified and protected from further abuse,” engineering lead Nikola Todorovic and product manager Abhi Chaudhuri wrote in a company blog post, September 3rd. “We’re making this available for free to NGOs and industry partners via our Content Safety API, a toolkit to increase the capacity to review content in a way that requires fewer people to be exposed to it.”

Internet Watch Foundation, a division focused on minimizing the online availability of child sex abuse images, appreciated the tool’s development, saying it will make the internet safer.

What are the plans to further develop Google’s new tool

The AI program is considered to be perfectable, so the team will continue working all the angles. The tool misclassified some images; however these same images were also misclassified by pathologists.

The program was, however, able to correctly assign cancer type to 45 of 54 images previously misclassified by at least one pathologist.

Even more impressive, it only took the model 20 seconds on average, running on a single-GPU PC, to calculate classification probabilities.

The team plans to keep training its AI program with data. The aim is to make it determine what genes present mutations in a given cancer with more than 90 percent accuracy. At that point, researchers intend to begin the process of obtaining government approval to use the technology clinically, in the diagnosis of several cancer types.