Artificial intelligence. Deep learning. Machine learning. No matter the name, the subject simultaneously creates excitement about the possibilities and begs questions about the unknown.
During a Monday morning presentation, two experts will discuss what the technology means for the future of urology. The 15-minute presentation, titled Deep Learning in Urologic Diagnostics: Prostate Histology, will begin at 7:45 am during the Next Frontier plenary presentation in Hall E at Moscone North.
Larry Goldenberg, MC, OBC, MD, FRCSC, professor of Urology at the University of British Columbia in Vancouver, Canada, will begin with a look at how deep learning can be applied to histopathology in diagnosing prostate cancer in the not-too-distant future.
“The development of high speed image processing software with modern digital scanners and ever increasing computation power is allowing us to train computers to perform very complex tasks that we typically do as specialists,” said Dr. Goldenberg, who is also Chair of the Canadian Men’s Health Foundation. “We’re able to digitize histopathology slides in high resolution, which gives us an opportunity to improve efficiency. Pathologists are quite burdened by the task of examining hundreds and hundreds of slides. By training a computer to differentiate cancer from non-cancer on a histology slide, we would enable the throughput of more and more slides.”
Dr. Goldenberg will share data from his work to train computers to not only identify cancer, but to grade cancer as well.
“We digitized a large cohort of 502 radical prostatectomy cases and we found correlation between the computer and pathologist to be excellent,” Dr. Goldenberg said. “The computer is able to identify the areas of cancer and also the Gleason Score. In our study the computer’s accuracy for differentiating cancer from benign was 90 percent, and for differentiating low to high grade classification it was 77 percent.”
Even with such positive results, Dr. Goldenberg said he does not believe computers will make pathologists or radiologists obsolete. Instead, computers will save specialists vast amounts of time by quickly reviewing large numbers of scans and alerting physicians to suspicious areas that need review, he said.
In another example of how machine learning may shape the future of urology diagnostics, Dr. Goldenberg said computers are being trained to predict outcomes.
“You can show the computer a slide of a patient with cancer and then, using an outcomes database, tell the computer what happens to that patient five or 10 years later,” Dr. Goldenberg said. “If you present the computer with thousands and thousands of cases with related clinical outcomes, unsupervised, deep learning algorithms will recognize features or patterns that will predict how patients will do with a certain disease. And it will be able to do so better than humans. I suspect the computer will be able to see things in a digitized pathology image or MRI [magnetic resonance imaging] that human occipital lobes cannot appreciate.”
The session’s other presenter, Peter Pinto, MD, Head of the Prostate Cancer Section and Director of the Fellowship Program in the Urologic Oncology Branch at the National Cancer Institute of the National Institutes of Health (NIH), will discuss why he believes deep learning will be the next innovation in medical imaging.
“The opportunity for machines to reference images is already beginning in medicine,” he said. “For example, in dermatology they’re training computers to recognize visual patterns to diagnose skin lesions. The computer can do this better and faster than the physician. This can help the physician be a better diagnostic clinician. This is where urology will benefit from deep learning — in diagnostic imaging.”
Dr. Pinto will focus his talk on the current needs in urological imaging.
“I believe the need right now is for MRI to detect prostate cancer,” he said. “Some of the work in our lab at the NIH and by others during the past 20 years has shown that with highly trained urology and radiology physicians, multiparametric prostate MRI has the ability to detect aggressive, high grade prostate cancer at an early stage. The difficulty in adopting this is in the need for all radiologists to be able to interpret the MRI correctly. At the NIH, we are looking at deep learning to improve how the computers can identify regions of cancer and bring the attention of the physician to those areas.”
Dr. Pinto and his colleagues are training deep learning computers with prostate MRI and prostatectomy histopathology correlation studies.
“Patients who enroll in our trials at the NIH undergo prostate MRI, and for those who go on to surgery, we section the prostate in a patient-specific, 3D printed pathology mold that correlates directly with our MR image slices,” he said. “We’re training computers to see the areas the MRI correctly identified and the areas that it missed. This is pattern recognition that human brains perform every day, but it’s remarkable how quickly the computers can process hundreds of thousands of images. The goal is not to replace radiologists. The goal is to help them.”
Dr. Pinto said the technology will be a new adjunct, like any other resource that physicians can use to become better diagnosticians.
“We’ve been a bit slow to incorporate this technology, but our trainees are ready,” he said. “They’ll easily incorporate this into practice. We may see deep learning in everyday use within the next 10 years.”