Google Makes Progress on Teaching Computers to Diagnose Cancer

Google machine learning research group are developing algorithms that automate the detection of breast cancer from large-scale slide images of tissue samples.

Healthcare IT

An ongoing Google investigation into the use of deep-learning techniques in the field of digital pathology shows early promise the company said this week.

Researchers from Google’s machine learning group are developing algorithms for the automated detection of breast cancer metastases from whole slide images of biological tissue samples.

Using images supplied by the Radboud University Medical Center in the Netherlands the researchers have been training the algorithms to automatically recognize signs that an individual’s breast cancer may have spread or metastasized to nearby lymph nodes.

Trained pathologists currently perform the task by manually reviewing very high-resolution, 10-gigapixel slide images of tissues samples from a patient. Each patient can sometimes have multiple slides each of which have to be magnified greatly and reviewed in minute detail before a diagnosis can be made, Google researchers Martin Stumpe and Lily Peng said in a blog this week.

Even with extensive training, pathologists can often arrive at different diagnoses from the same set of images. In fact, agreement between pathologists for some types of breast cancer diagnosis is just 48 percent, which isn't surprising given the sheer amount of data that needs to be reviewed, Stumpe and Peng said.

Google’s research shows that deep learning techniques could play a useful role in helping pathologists with their diagnoses, the two researchers said. Earlier research that Google has done in this area with a project called Inception showed that deep learning approaches worked reasonably well in cancer diagnoses.

With additional work on customizing algorithms and figuring out ways to examine slide images at different magnifications Google researchers have shown it is possible to train a model that matches or even exceeds the performance of a trained pathologist, the two Google researchers claimed.

So-called prediction heat maps generated by Google’s deep learning algorithm in a biomedical imaging contest last year achieved a higher-score on a certain performance metric for detecting cancer, than that achieved by a pathologist who spent 30 hours on the same task, they said.

“Even more exciting for us was that our model generalized very well, even to images that were acquired from a different hospital using different scanners,” Peng and Stumpe said.

Deep learning is more of a focused sub-discipline of machine learning. It focuses on specific machine learning tools to develop methods for addressing real-word problems. Google’s Automatic Speech Recognition technology is one example of an app powered by deep learning techniques.

While the early results of the work Google has done in the area of clinical pathology are promising, a lot more work remains to be done, Stumpe and Peng noted.

The deep algorithms have proved to be capable at tasks for which they are specifically trained, but lack the broader knowledge and experience that a trained pathologist would bring to task.

So, while the algorithms might work at detecting breast cancer metastasis they would be less useful in detecting abnormalities for which they had not been trained like detecting inflammation. The metric, which showed Google’s deep learning algorithms performing better than a pathologist, is also not truly reflective of real-life diagnosis making.

“To ensure the best clinical outcome for patients, these algorithms need to be incorporated in a way that complements the pathologist’s workflow,” the Google researchers said. If done correctly, they could help improve efficiency for pathologists, they added.

Jaikumar Vijayan

Jaikumar Vijayan

Vijayan is an award-winning independent journalist and tech content creation specialist covering data security and privacy, business intelligence, big data and data analytics.