Researchers using artificial intelligence to improve the diagnosis and management of incidental kidney tumors
About 70,000 new kidney tumors are diagnosed in the United States every year, most of them found incidentally on a CT scan or an MRI. Most are small, <4 cm in diameter, and there is no reliable way to predict with imaging alone whether a kidney mass is cancer or not. Unlike other tumors that are amenable to biopsy, renal mass biopsy has a fairly high nondiagnostic rate.
"The reason it matters is that up to 20% of these tumors can be noncancerous," explained Russell S. Terry, Jr. M.D., an assistant professor and director of minimally invasive surgery education and new technologies in UF College of Medicine's department of urology. "If we remove people’s kidneys or part of their kidneys for 100% of these small masses, then up to 20% of the surgeries might be unnecessary."
Urologists and radiologists have investigated nonsurgical methods to determine which kidney tumors are cancerous, mainly by looking at the subtleties of how they appear on CT scans.
"Previous efforts to do this have been limited by our inability to process and compare large numbers of scans to find patterns, often subtle patterns," Terry said. "Even if we find the patterns, it is difficult to then validate them on a large enough scale to make sure they are accurate across the general patient population. So surgically removing the mass remains the only way we can definitively establish whether or not it is cancer, and that is a major surgery."
Terry is working to develop a computer algorithm that will automate the review of kidney masses on CT scans.
"The University of Florida and the NVIDIA Corporation recently made an unprecedented investment in the field of artificial intelligence," he reported. "Through their public–private partnership, UF now houses the world’s fastest AI supercomputer in higher education."
One form of AI is deep learning, a form of machine learning in which computer systems learn to recognize patterns within layers of complex data, like medical images, and correlate the patterns with outcomes. Once the system understands this relationship, it can then be used to predict outcomes, like whether or not a mass might be cancer. Terry is partnering with one of UF’s experts in the division of cancer informatics, Jiang Bian, Ph.D., an associate professor in the UF College of Medicine, and several of his graduate students to improve AI recognition of kidney cancer.
To solve the problem of evaluating kidney masses, the first step is to teach a computer to be able to find a kidney.
"That is called registration. Once the computer can register the kidney, you have to teach it what is normal and what is abnormal — you have to teach it to find a mass within a kidney. That is the second step of the registration," Terry explained.
Other groups have already developed computer algorithms that can do those first two steps.
"So, we have a bit of a head start," Terry noted. "What has not been described is the difficult and important step of correlating subtle findings — comparing what these masses look like with the final pathology results, and doing that enough times that the computer learns that pattern A is associated with pathology A, pattern B is associated with pathology B and so forth. That is what we are hoping to do.
"The goal is to be able to analyze a patient’s CT scan with an algorithm that can accurately predict whether or not their kidney tumor is likely to be a cancer. That way, their treatment and management can be tailored more precisely to them," Terry said. "There’s a lot of potential here for us to be smarter about how we diagnose and treat incidental small renal masses. I believe this is only the beginning of how you’ll see urologists use AI to improve the care of our patients."