When a doctor orders a biopsy for a patient whose clinical findings raise the possibility of prostate cancer, it generates dozens of slides for a pathologist to review.
“There are 12 to 18 parts that are tested, and we receive three slides per part,” says Joshua Warrick, MD, professor of pathology and director of genitourinary pathology at Yale School of Medicine. “So, you’re often reviewing 50 or more slides for a patient.”
The genitourinary subspecialists of Yale Pathology closely inspect each slide. If they detect cancer, they provide precise grading and measurement, which distinguishes aggressive tumors from slow-growing ones and informs the urologist’s treatment. In addition to using a microscope, Yale pathologists use several technologies to help with their diagnoses. These include immunohistochemical analysis, whole-slide scanners, and automated tissue processors.
They recently added to their toolbox an AI platform that performs a quality-assurance check by assessing biopsy slides with a tool trained on a massive database of prostate biopsies. Yale uses the platform to compare the digitized slides the pathologist reviews against a database of tens of thousands of slides of prostate cancer – providing an instant quality-control check. The Paige AI platform has FDA clearance for use as a quality-assurance tool.
“It’s like a safety net to catch small things a pathologist may miss,” Warrick says. “It’s designed to catch cancer but not necessarily be accurate. Diagnosis remains with the pathologist.”
As he finalizes his report, Warrick compares it to the AI review. If the AI flags as concerning a slide he called benign, he will take another look. “It’s usually a tiny, atypical gland or two that the AI catches and we miss,” he says. “It’s not sufficient to diagnose cancer, but important to note for future follow up.”
The new platform is a deep learning neural network, a subset of AI modeled after the human brain. It has many layers of interconnected algorithms that process data and learn complex patterns to perform the review. The platform is Yale Pathology’s latest embrace of AI to support the work of its pathologists. Yale joins a handful of academic centers that are using the prostate AI platform.
"Digital pathology is transforming how we deliver diagnostics, advance translational research, and integrate AI into clinical care,” says Chen Liu, MD, PhD, Anthony N. Brady Professor of Pathology, chair of pathology, and chief of pathology, Yale New Haven Hospital.
Once someone at Yale Pathology scans a slide into the current Yale system, a bridge – built by the Pathology Informatics Team – carries the data to the new AI platform. It runs the algorithm, and the results are available at the time the pathologist reviews the slides. Joseph Celano, senior software engineer, made a major contribution to building the system, and Kevin Chieppo, software engineer, also contributed significantly. The entire informatics team participated throughout the project, according to Peter Gershkovich, MD, MHA, associate professor of pathology and informatics team leader.
“We’re still collecting metrics and other data to see how impactful this (new platform) is in delivering cutting-edge patient care within our health system, but AI-assisted diagnosis is the direction the field is headed,” says Sudhir Perincheri, MD, PhD, MBBS, associate professor of pathology. Perincheri led one of the first published studies that demonstrated the promise of using the Paige AI prostate algorithm.
If pathologists determine the new platform is consistently accurate in identifying cases where there is no cancer, they could conceivably use it as a prescreening tool, allowing them to focus on cases where cancer is more likely to exist.
“It's early and we are still figuring out the optimal way to use the system,” Perincheri says. “I think the jury is still out on how the technology will evolve to impact clinical workflows in pathology. Ultimately, AI will help automate and standardize large aspects of the workflow, help triage cases, extract metrics from histology images tied to patient outcomes, and deliver superior, patient-centered care tailored to each patient.”