Deep learning, a form of artificial intelligence (AI) where computers are taught to recognize patterns in huge datasets, can now be useful in identifying breast cancer.

Google reported March 3, Friday, that it has achieved groundbreaking results in using AI to analyze thousands of cancer cell slides from a Dutch university and diagnose the common form of cancer.

How Deep Learning Works

A pathologist’s report is usually the gold standard in diagnosing many diseases, including breast cancer. But even with extensive training and experience, there can be variations in diagnoses by different pathologists.

“[A]greement in diagnosis for some forms of breast cancer can be as low as 48 [percent], and similarly low for prostate cancer,” Google wrote in a blog post. “The lack of agreement is not surprising given the massive amount of information that must be reviewed in order to make an accurate diagnosis.”

Using images provided by the Radboud University Medical Center, Google trained algorithms using deep-learning approaches such as Inception — or GoogLeNet — to match or exceed a pathologist’s performance. After being pitted against an experienced pathologist to examine slides in an unlimited time, GoogLeNet scored 89 percent in accuracy while the human achieved 73 percent.

The technology, though, is not designed to replace human specialists.

“What we've trained is just a little sliver of software that helps with one part of a very complex series of tasks," Lily Peng, the project manager behind Google's work, told CNN.

Peng explained that given its extreme sensitivity, their AI system can flag things that humans will miss. However, it can also provide a false positive that a human pathologist can confirm isn’t cancer at all.

The algorithm can localize and find the tumors while the doctor can provide the ultimate verdict that there’s no cancer, Peng added. The model, for instance, has not been trained to classify certain things including inflammatory process, autoimmune disease, or other kinds of cancer.

Google’s new work is still in research phase, but the company thinks it’s “off to a very promising start” and will accelerate progress in the space. Jeroen van der Laak of the Dutch medical institution also believed that the first cancer algorithms can be available in a few years’ time, and large-scale standard use will take place in around five years.

Breast Cancer Diagnosis

Breast cancer remains the most pervasive type of cancer in women worldwide, with data from CDC showing that about 230,000 cases are diagnosed in women every year.

A study last month showed that the trauma of getting a false-positive result from one’s mammogram can cause many women to delay or skip the next screening. False positives occur when an aberration on a mammogram reflects cancer, but additional tests like biopsy or added imaging would later show as benign.

In the study, women who had false positives on their first mammogram delayed their second screening by 13 months on average. Those with false positives were likely to be on the younger side, premenopausal, getting their first screening, black, and with denser breasts.

A separate study looked at cell shapes in millions of imaging scans of over 300,000 breast cancer cells, and discovered that changes in shape — resulting from physical pressures on the tumor — translated into changes in gene activity, which were tied to actual clinical disease outcomes.

This, according to experts, could enable healthcare providers to tell based on the cells’ appearance how aggressive the cancer is and how rapidly it is likely to spread.

ⓒ 2021 All rights reserved. Do not reproduce without permission.