Peering into the microscope to sift through millions of cells to spot just a few cancerous ones can be very labor-intensive with conventional methods. The new AI system is able to tackle this task quite well, the researchers found.
The artificial intelligence (AI) system is "based on deep learning, a machine-learning algorithm used for a range of applications including speech recognition and image recognition," explains Andrew Beck, an associate professor in pathology at Harvard Medical School, who heads the team developing the new system at Beth Israel Deaconess Medical Center (BIDMC), in Boston, MA.
Prof. Beck and colleagues demonstrated the new AI system in a competition held at the annual meeting of the International Symposium of Biomedical Imaging (ISBI 2016) in Prague in April.
He and his colleagues are developing AI methods that train computers to interpret pathology images to improve the accuracy of diagnoses.
The approach they are using teaches computers to interpret the complex patterns seen in such images by "building multi-layer artificial neural networks," says Prof. Beck.
The process is thought to be similar to the way learning takes place in the layers of neurons in the neocortex of the brain, the region where thinking occurs.
The team put the new AI system to the test at the ISBI 2016 meeting by getting it to examine images of lymph nodes to decide whether or not they showed evidence of breast cancer.
The team started training the AI system with hundreds of training slides labeled by pathologists to show the difference between cancerous and normal cells.
They then extracted millions of the training examples and used deep learning to build a model to classify them. This included identifying each time the AI system got it wrong and then re-training it using more and more of the difficult examples.
The test at the meeting showed the AI system on its own correctly diagnosed the presence of cancer 92 percent of the time, just 4 points short of the 96 percent accuracy achieved by a human pathologist.
"But the truly exciting thing was when we combined the pathologist's analysis with our automated computational diagnostic method, the result improved to 99.5 percent accuracy," notes Prof. Beck. "Combining these two methods yielded a major reduction in errors."
Prof. Beck explains that pathologists have been working on using digitized images and machine learning to improve and speed up diagnosis for decades, but it is only recent improvements in scanning, storage, processing, and algorithms that are making it possible to make significant progress.
He says the results in the ISBI competition show that what the AI system is doing is "genuinely intelligent" and, when you combine it with human ability, it will lead to more precise and clinically valuable diagnoses.
One of the competition organizers, Dr. Jeroen van der Laak, who leads a digital pathology group at Radboud University Medical Center in the Netherlands, says the results clearly show that AI is going to shape the way pathologists use images in the future.
"Identifying the presence or absence of metastatic cancer in a patient's lymph nodes is a routine and critically important task for pathologists. Peering into the microscope to sift through millions of normal cells to identify just a few malignant cells can prove extremely laborious using conventional methods. We thought this was a task that the computer could be quite good at - and that proved to be the case."
Prof. Andrew Beck
The team is publishing a technical report on the new AI system in the open access arXiv.org repository.