Machine learning software may help pathologists make earlier and more accurate diagnoses
In Boston, two major academic centers are teaming up to apply big data and machine learning to the problem of diagnosing cancers earlier and with more accuracy. It is research that might have major implications for the anatomic pathology profession.
A collaborative effort between teams at Beth Israel Deaconess Medical Center (BIDMC) and Harvard Medical School (HMS) has resulted in an innovation that could result in more accurate diagnoses in the pathology laboratory. The teams have been working on a machine learning software program that will eventually function as an artificial intelligence (AI) to improve the accuracy of diagnostics. They hope to someday build AI-powered computer systems that can accurately and quickly interpret pathology images.
Using AI to Diagnose Disease
The teams are led by Andrew Beck, MD, PhD, Director of Bioinformatics at the Cancer Research Institute at BIDMC, Associate Professor at Harvard, and co-founder of PathAI, a pathology technology company focused on developing tools that will use AI in the clinical diagnosis of cancer and other diseases.
“This approach teaches machines to interpret the complex patterns and structure observed in real-life data by building multi-layer artificial neural networks in a process which is thought to show similarities with the learning process that occurs in layers of neurons in the brain’s neocortex, the region where thinking occurs,” Beck said in a BIDMC news release.
Of course, AI has been the goal of many researchers across a broad range of specialty areas, so it is not at all surprising that the methods are being developed for use in the pathology laboratory. A paper published in 2014 on BioMed Central written by Kenneth Foster, PhD, of the Department of Bioengineering at the University of Pennsylvania, et al, asserts, “In recent years, there has been a dramatic increase in the use of computation-intensive methods … under the rubrics of artificial intelligence or machine learning.” The paper was published on BioMedical Engineering OnLine and goes on to discuss “some major pitfalls of the technique.”
Using AI to Speed Up Laboratory Operations
The collaborative research team from BIDMC and HMS had the opportunity to demonstrate their work during a competition at the International Symposium of Biomedical Imaging (ISBI) recently. They placed first in two separate categories. The competition involved finding breast cancer in samples of lymph node tissue.
The competition represented a task that is not uncommon in the pathology laboratory. “Identifying the presence or absence of metastatic cancer in a patient’s lymph nodes is a routine and critically important task for pathologists,” said Beck, adding that it is also “extremely laborious using conventional methods.” Using the AI could cut down on the repetitive and time-consuming work, freeing the pathologist to do other tasks.
The machine learning method involved training the computer to tell the difference between cancerous tumor cells and normal cells through a multi-layered process. Dayong Wang, PhD, Research Fellow at Harvard Medical School and member of the research team, explained that they began the machine learning process by using training slides that had been labeled cancerous or normal by a pathologist, adding, “We then extracted millions of these small training examples and used deep learning to build a computational model to classify them.”
The next layer of the process involved identifying particular samples that are likely to result in mistakes, and including more of them in the computer’s learning process. Beck said that although this sort of research has been pursued for decades, “It’s only recently that improved scanning, storage, processing, and algorithms have made it possible to pursue this mission effectively.
“There have been many reasons to think that digitizing images and using machine learning could help pathologists be faster, more accurate, and make more accurate diagnoses for patients,” Beck concluded.
Jeroen van der Laak, PhD, Assistant Professor at Radboud University Medical Center in the Netherlands, helped to organize the ISBI competition. He was surprised by the results, saying, “The fact that computers had almost comparable performance to humans is way beyond what I had anticipated.”
AI Accuracy Nearly That of Human Pathologists
Such has not always been the case for machine learning software. Dr. Tatjana Zrimec, Senior Lecturer at University of New South Wales and head of the Master of Health Informatics post-graduate program at the University of New South Hampton, and Igor Kononenko, PhD, Head of Laboratory for Cognitive Modeling, Department of Artificial Intelligence at the University of Ljubljana in Slovenia, said, “In spite of huge development of the biomedical technology, the diagnostic accuracy is in many cases rather low,” in a paper titled “Feasibility Analysis of Machine Learning in Medical Diagnosis from Aura Images.”
One of the important aspects of this technology is that it is designed to complement the work that medical pathologists do. This particular machine learning system can be used to diagnose cancer earlier and more accurately, which is beneficial to patients, physicians, and pathologists.
In the ISBI competition, the AI method had an accuracy rate of about 92%. Aditya Khosla, PhD, co-founder and CTO at PathAI, and a research assistant at the MIT Computer Science and Artificial Intelligence Laboratory, who participated on the winning team, says the results “nearly matched the success rate of a human pathologist, whose results were 96% accurate.”
But it wasn’t just the high accuracy rate of the AI method that excited Beck. It was the combination of human intelligence and machine learning. “The truly exciting thing was, when we combined the pathologist’s analysis with our automated computational diagnostic method, the result improved to 99.5% accuracy. Combining these two methods yielded a major reduction in errors.”
In cases where determining the severity of the disease is important—such as with lung cancer—a machine learning tool could be critical. An article in Healthcare Informatics describing a similar technology being developed to help diagnose lung cancer stated that the classification system used to determine the grade and stage of cancer doesn’t work well for lung cancer. Michael Snyder, MD, FACS, Professor and Chair of Genetics at Stanford University, who led the research said, “Ultimately this technique will give us insight into the molecular mechanisms of cancer by connecting important pathological features with outcome data.”