Doctors Peter Chang and Daniel Chow are co-directors of the UCI Center for Artificial Intelligence for Diagnostic Medicine (CAIDM)
November 25, 2019 –
Doctors Peter Chang and Daniel Chow have landed in the right place at the right time. About a year ago, the two neuroradiologists at UC Irvine Medical Center launched the UCI Center for Artificial Intelligence for Diagnostic Medicine (CAIDM) on the fourth floor of CALIT2. It’s an ideal location for a center that needs to rapidly and securely access huge amounts of scientific data from multiple institutions.
The center focuses on developing and applying AI tools that can detect, characterize and provide prognosis for a variety of conditions in an effort to advance patient care, improve health outcomes and lower costs in all areas of health care. The burgeoning center integrates the expertise of health care professionals, scientists, software engineers and data scientists to provide a central research core for all UCI faculty, physicians and researchers.
The CALIT2 building is equipped with leading edge technologies to support data sciences, including access to the Pacific Research Platform (PRP), a high-capacity fiber optic network connecting data servers at all the UCs and their six medical centers, plus several large research institutions throughout the country.
This access combined with today’s advancements in deep learning technology are converging in a way that holds great promise for using artificial intelligence to improve medical care, especially in radiology. Chow and Chang are capitalizing on this opportunity.
In fact, it was while working a shift in the medical center’s imaging services department that Chow thought there had to be a more efficient way to help patients who are waiting in the emergency room for results of their imaging tests (X-rays, CT scans, MRI scans). He had an idea for applying AI technology to the multiple brain scans that are taken for each patient who comes to the ER with a headache, and whose images must be methodically reviewed by a radiologist. Chow says the clinicians have to rule out brain bleed, hemorrhage or any sign of stroke. Because if a hemorrhage is identified, a neurologist or neurosurgeon must be called in right away to assess the course of treatment.
“Ninety percent of the scans obtained in the ER are going to be normal, but we have a ton of images to go through and we go in order. It would be nice if something told us, ‘look at me right away,’” says Chow, assistant professor in residence with an appointment in radiology and neurology.
That was the first AI tool that Chow and Chang developed and it’s currently in clinical trials at the medical center. They designed a custom algorithm that identifies and assesses a hemorrhage at the point the scan is captured, then immediately sends an alert to a specially set up inbox. This tells the doctor to review those flagged scans right away. “We’re not replacing the radiologist, just helping speed up the process for results of concern,” explains Chow.
Chow and Chang completed their fellowship training together at UC San Francisco. Chang is both a radiologist and self-taught software engineer, so he brings insights in both domains. He had been consulting with AI startups in health care institutions while finishing his medical degree when Chow called him to join the effort at UCI. He saw CAIDM as the perfect chance to start a center from the ground up.
A year later, the center has nine permanent staff, half a dozen visiting researchers, and multiple residents, fellows and medical students. And there are more than a dozen projects in the works. In addition to the brain hemorrhage triage tool, they are developing several AI algorithms capable of identifying normal vs. abnormal anatomy of the head and neck. They have AI methods in the works for automated detection of prostate lesions and kidney cancer as a clinically efficient diagnostic tool. And there are a couple projects involving breast imaging: one that enhances detection of breast lesions on MRIs and one that automatically segments fibro-glandular from surrounding breast tissue, which will help clinicians better evaluate a patient’s risk of possible breast cancer. In collaboration with other physicians interested in applying AI to their practices, the center is working on tools that can detect lung nodules on chest CTs, assess the risk of prostate cancers and do a virtual brain biopsy.
Although artificial intelligence in medicine is growing, it’s still an emerging area, with uncharted territory. Chang explains the advantage they have at the UCI center. Both he and Chow are practicing clinicians. “Most of the innovation in machine learning in medicine has been driven by Ph.D.s, not M.D.s,” says Chang, “where the computer scientists look to take the techniques that they’ve learned in self-driving cars and facial recognition and apply it to medicine.”
Instead, Chang and Chow are bringing clinicians to the table and letting them determine what they want to study and what problems they need to solve, and then the two work together to design an AI application and determine how best to incorporate it into the clinical setting. The algorithm has to match the application, whether it is to make a diagnosis, a detection, a quantification or a comparison. And an algorithm is only as smart as the data it’s trained on.
This brings us back to the right place at the right time. “With all these research institutions connected through the PRP, it’s an amazing resource, like a distributed supercomputer,” says Chang. And with sensitive patient data, without the PRP, it would be almost impossible to share outside of the hospital firewalls. “In the world of medical AI, there is a real bottleneck when it comes to data.”
“I am excited about Peter and Daniel’s idea of using the PRP to couple the training of machine learning on radiological images at two different sites (UCI and UCSD), without having to move the data out of its protected environment,” said Larry Smarr, CALIT2 founding director and principle investigator of the PRP. “This is a novel use of the PRP that we had not imagined when we wrote the proposal to NSF five years ago.”
Chang explained, “An algorithm that is trained at one hospital would not be able to generalize and work in another hospital with the same degree of accuracy. Every hospital is different — the imaging technology, the patient population, the configuration of patient care, etc. But by training an algorithm on multiple hospitals, we end up with a more powerful and general tool.”
CAIDM has built a full enterprise level infrastructure to provide de-identification imaging services to all UCI researchers in a secure process to help streamline the current workflow. They are able to query, download, anonymize, store and transmit data for optimal efficiency across high bandwidth data center networks.
As a result, CAIDM’s algorithms are extra smart because they have seen data from so many institutions. “It’s a diverse representative sample because we’ve found a way to share without compromising HIPA rules,” says Chang.
Both Chang and Chow are excited about the potential collaborations of being on the campus as opposed to the medical center. Chang, assistant professor in residence in radiology and computer science, leads the new health care AI curriculum, which trains the next generation of physician-scientists in understanding and developing cutting-edge AI tools. His students are able to propose and design projects, and this contributes to a community of ideas, in which the center is educating and executing.
Chow says that the idea of personalized or precision medicine is not far from reality. “I think we are on the cusp. With this AI technology, we can take on questions we never thought of tackling before; we can really push the boundaries.”
– Lori Brandt
