The method for diagnosing breast cancer has gone mostly unchanged since the 1920s: doctors examine a small set of cancer cells to determine the disease’s aggressiveness and the patient’s prognosis.
This week, however, computer scientists and pathologists at Stanford University described how they trained a computer system to analyze tissue samples on slides and diagnose breast cancer more accurately than a doctor.
In a paper published in Science Translational Medicine magazine , the computer scientists at the Stanford School of Engineering and pathologists at the Stanford School of Medicine explain how they taught a computer to learn the signs of cancer.
Andrew Beck, a doctoral candidate in biomedical informatics and the paper’s first author, said he hopes the technology could be used in healthcare facilities to diagnose breast cancer in about five years.
“The next step we’re working on is using whole-slide images,” Beck said. “This [test] was all done on various small images that are easier to use in a research setting. There’s work to do on inter-institutional variability. There’s a lot of quality control testing that will need to be done to ensure it’s providing good data from different institutions.”
Beck said the computer program, called Computational Pathologist, or C-Path, runs on a common x86 computer with 16MB of memory. It automatically analyzes images of cancerous tissues and then predicts the patient’s chance of survival.
Breast cancer tumors are graded on a scale of 1 to 3, with grade 1 being a slow cell growth rate and 3 being the fastest growth rate.
A pathologist looks at the tumor cells and checks for three microscopic features: the amount of cancer made up by tubular structures; the speed of cell division; and the size of the cell’s nucleus and uniformity. C-Path does the same thing, only with a higher degree of accuracy, according to Beck.
Currently, however, the slides being evaluated by C-Path are only .5 millimeters in diameter. Each slide takes 10 minutes to read. A typical biopsy slide is 3 centimeters by 2 centimeters in size.
“Scaling it up is an issue. There are like 1,000 of these little tiny images in a full slide image. So you can’t take 10 minutes on each of those,” Beck said. “So we’re working on doing more in parallel and seeing what parts of the [analysis] need to be done in high resolution.”
The researchers first had to train C-Path using existing tissue samples taken from patients with a known prognosis. The computers were taught to look at the same three features pathologists study. The computers then compared the analyses against the results of the known cancer patient data.
Over time, the computers evolved their models to better determine what features of the cancers matter most and which matter less in predicting patient survival, according to Daphne Koller, a professor of computer science and senior author of the paper.
C-Path assesses a total of 6,642 cellular factors. “These measurements were used to construct a prognostic model,” the paper states.
After learning how to evaluate the microscopic slides, C-Path was then shown new tissue samples from cancer patients it had not already checked. The computer’s results were checked against the patient’s known outcomes. The C-Path computer had a statistically higher level of accuracy compared to physician-based evaluations of those same patients.
“This will be a tool to help guide physicians,” Beck said. “It won’t replace them…. But this could be a decision support tool they use in their overall interpretation.”
