By BECKY GILLETTE
Artificial intelligence has been in the news a great deal recently with Congressional hearings where some speakers advocated that Congress establish licensing and guidelines to mitigate the risk of harm that some see possible with the rapid development in AI. Licenses could be revoked if AI is used improperly.
But machine learning (ML), a subfield of AI that uses algorithms to produce models that can perform a variety of complex tasks, has broad promise in medicine as a tool aimed at providing more accurate cancer screening information. That gives the potential to save time, improve treatments and potentially help reduce costs and patient stress, said Fred Prior, PhD, an AI expert who is head of the Department of Biomedical Informatics at the University of Arkansas for Medical Sciences (UAMS).
“There are many aspects to this,” Prior said. “Just as we are seeing with automobiles or home appliances, machines used in medicine are getting smarter. The tools physicians use, whether CT scanners or patient monitors, are going to have AI algorithms built into them that will improve workflow and reduce errors. It will be easier to filter out things that don’t need the physicians’ attention. Now there are a lot of false alarms and alerts that tend to make people turn off those functions rather than use them to filter and get the information they need.”
Prior said researchers are finding ML tremendously useful to explore new research questions that previously couldn’t be tackled.
“Algorithms are given huge amounts of data that allow them to learn from experience which provides the ability to do a very good job of understanding the problem,” Prior said. “We are doing a lot of work in the AI area trying to better understand what the algorithms can teach us about cancer.”
Deep learning algorithms use multiple layers to progressively extract higher-level information from raw data, and have potential for helping learn new things about cancer.
Generative modeling for cancer detection is a relatively new idea. Prior said algorithms can be built to learn the characteristics of a mammogram and then those models are used to create new images of people who don’t exist. Prior said a generated image or synthetic data can be used for testing without exposing the real person to more scans.
“One of the problems with trying to diagnose breast cancer is the features you are looking at are relatively small,” Prior said. “If radiographic density is high, it can be harder to detect. UAMS is currently working with researchers all over Europe with an AI in cancer imaging project funded by the European Union to develop synthetic data that can be used to train machine learning algorithms by augmenting training data with hard-to-find cases. UAMS is the only institution in the U.S. involved in this study. It is a way of expanding our training information through algorithms. It turns out to be extremely useful to train algorithms that can generalize and deal with the high variability within patient populations.”
Prior said they need millions of cases in order to represent the variance in the human population. That is only possible with AI.
AI also has great potential for lung cancer screening. A National Cancer Institute trial has established the best type of CT for lung cancer. But in that trial, the number of false positives was very high. The trend is to err on the side of false positives in order to protect patients. But it is highly stressful for patients to be told they have cancer when they don’t.
“What we are doing is building a machine learning algorithm to drastically reduce false positive rates,” Prior said. “We want to reach that balance where we have a machine with the potential to provide fewer false positive and false negative rates than a radiologist. I’m not trying to replace radiologists. But if I can take out 80 percent of the easy cases and get them right, then the radiologist’s workload is greatly reduced making screening more accessible and economically viable.”
Prior said if they can improve screening with ML that is as good or better than a radiologist, it makes it more economically viable to have widespread screening.
“We have published algorithms that show we can do as well as radiologists to reduce false positives but not necessarily false negatives,” he said. “That is the harder one. That is what we are working on now. It is very significant to improve screening and reduce false positives where patients have a few horrible weeks thinking they have cancer when they don’t. But we don’t want to miss anything.”
Prior said liquid biopsies that test blood instead of tissue for cancer are a brilliant idea.
“It minimizes the stress on patients because you can do a blood sample instead of a tissue sample,” he said. “Combining a blood draw with imaging could be very helpful but the research process is just getting started.”
Prior serves as principal investigator and director of the National Cancer Institute's Cancer Imaging Archive project and is the lead PI of an NCI ITCR team exploring the integration of radiomics and pathomics.
For more information, see the UAMS Creativity Hub for Artificial Intelligence in Health (CHAI) https://medicine.uams.edu/research/hubs/artificial-intelligence-for-health/ which has a mission of “ensuring that diverse voices define the future of how AI and ML are changing our world. Some argue that we are entering an age of better, more targeted healthcare. Others warn we may inadvertently reinforce existing health disparities. Undoubtedly, algorithms have great potential to identify patterns in complex data and revolutionize medicine. Let’s have all voices at the table.”
