Stanford Study Will Inspire Any IT Pro Intrigued By Machine Learning

13 Ways Machine Learning Can Steer You Wrong

For IT organizations, machine learning is looking like an essential capability in the decade ahead. For the past few months, Google CEO Sundar Pichai has been extolling the value of AI and machine learning to his company. Gartner has added machine learning to its 2016 Hype Cycle, putting it at the peak of inflated expectations.

The Hype Cycle, said Gartner research director Mike J. Walker in a statement, lists technologies that show "promise in delivering a high degree of competitive advantage over the next five to 10 years."

Now, researchers at the Stanford University School of Medicine have demonstrated trained computers can outperform doctors when evaluating the slides of lung cancer patients, a finding which underscores the value of machine learning for data analysis tasks involving image recognition.

The Stanford findings underscore the promise of machine learning in enterprises of all stripes.

In a paper published in Nature Communications last week, a group of scientists led by Kun-Hsing Yu described how software-based machine learning is "superior to the current practice utilized by pathologists who assess the images in terms of tumour grade and stage."

Traditionally, pathologists have evaluated patient tissue on glass slides through a microscope to determine the grade or severity of cancerous tumors and the stage to which the disease has progressed. But Michael Snyder, professor and chair of the department of genetics at Stanford, told the Stanford Medicine news service such assessments tend to vary among pathologists, who agree only about 60% of the time.

Computer-driven analysis of medical slides has been a challenge due to the large amount of information contained in whole-slide pathology. "The huge dimension of the original images made it extremely difficult to manipulate, and informatics workflows requiring manual tumour tissue segmentation were not feasible for millions of image tiles," the authors explain in their paper.

The Stanford researchers managed to create an informatics workflow that can be automated and scaled to address the analysis of data-dense imagery. The relevant details of tissue imagery on slides can be difficult to identify by manual inspection, the authors observe, but computers can be trained to identify salient features. The authors believe their approach can quickly and objectively assess the survival odds of lung cancer patients and lead to better therapeutic decision-making.

The researchers based their work on 2,186 images from the Cancer Genome Atlas, a national database of cancer data that includes information about the grade and stage assigned to the cancer imagery. They used those images to train their software to identify 9,879 distinct image characteristics, more than any person would weigh.

In an email, Snyder said the researchers' system performed 15% better on prognosis than pathologists.

[How is Intel approaching AI? Read Intel Buys Deep Learning Startup Nervana to Bolster AI Business.]

The diagnostic capability of computers goes back decades, but it hasn't always been as capable. In 1998, an expert system outperformed human physicians when all parties had the same information. It made accurate diagnoses 65% of the time, compared to 54% among participating doctors, according to a study titled "Comparison Between Diagnoses of Human Experts and a Neurotologic Expert System."

However, when the physicians had access to the complete medical records of patients, they surpassed the expert system, making accurate diagnoses 69% of the time.