Ulas Bagci, PhD, Professor of Computer Science at the Center for Research in Computer Vision at UCF talks about an AI system to detect cancerous tumors.
Interview conducted by Ivanhoe Broadcast News in May 2019.
I want to start by asking you how did the idea of this AI system come about.
BAGCI: Actually we were working this. I personally have been working on this more than 10 years now. And the first five years it was at NIH because I’m a former NIH scientist. And then I moved here and we continued this research. Lung cancer detection, the algorithms have tried to be addressed by several scientists. So we are one of those groups and one of the known groups in this field. We spent lots of years, lots of efforts. There were novel algorithms to solve this problem. So it’s coming up with potential solutions for helping radiologists and patients to detect lung cancer at early phase.
Why lung cancer in particular? How did you get started on that?
BAGCI: As a group we are working on several different diseases but lung cancer is one of our first choices to work because lung cancer is the number one cancer related death in the world. And in the United States it has pretty low survival rates. And any attempt to improve early detection and diagnosis is going to help several millions of people to live longer. And the development of new treatment strategies.
You were talking about helping to solve the problem that radiologists have. Could you go into a little bit about the problem that you’re helping them solve?
BAGCI: There are lung cancer screening procedures in the hospital. The patients undergo to a low dose computed tomography scanning. What radiologists are doing is they are looking at the images slice by slice because these are volumetric images. They look at the images by their eye. It’s a very time consuming procedure and they try to find if there is any abnormality ongoing in the lung. And some of these nodules, or pathology, are very small and it is very easy for a radiologist to miss it sometimes. Also they may see it but they still may not recognize whether they are significant or not. Sometimes it is large, but you may still miss it. And there may be other reasons (perceptual errors). Another reason is that sometimes it is missed by some other pathologies or normal tissues (camouflaged). So we are developing algorithms to search these pathologies on the scans and find it and try to present them to the radiologists for the final decision. We are actually helping them to improve their screening strategies and especially for very small nodules. The radiologists tend to miss them because they look like the normal tissues and it is very natural to miss those tumors.
And what’s the science behind getting a computer system to detect a tumor versus regular tissue?
BAGCI: We have images with millimetric resolution. We can capture millimetric tumors. But usually in the lung cancer screening strategies, if the tumor is less than 5-6 millimeters it is considered a very small tumor and a follow-up decision is taken by doctors. So our technique is actually very sensitive to capture. Even in that range, up to 6 millimeters. So those tumors are called early phase because they’re small. If you detect them then it is highly likely that the patient can be survived in the easier treatment strategies. And we are able to capture those tumors. Usually those are missed very easily.
What’s the percentage of you guys catching those tumors with this system versus a regular radiologist?
BAGCI: With these small nodules, it can be missed up to 30 to 40 percentage. Our overall system is more than 90 percent accurate, but for larger tumor size, our system is able to detect more than 95 percent as well. And we have another step after detection. So once you detect the lung nodules there is another step. We call it automatic diagnosis. We are developing artificial intelligence algorithms to tell you if inaudible it is malignant or benign. This is happening just after detection and we have 95 percent accuracy for diagnosing the detected nodules.
How are you training the computer systems to detect these differences?
BAGCI: We are collaborating mainly with NIH and the University of Pennsylvania and other institutes. We are collecting the data from these institutes and also doing a clinical trial in the United States. We obtained publicly available tomography images from clinical trials to do image analysis and once we actually collect these images, we have ground truths through either radiologist confirmed or the confirmed specimen. So we know where they are or their types (malignant vs benign). We feed that information into the computer and our AI algorithms are trying to find a mapping. It is very complicated mathematical equations. We are feeding these images and letting the algorithm to find this complicated mapping instead of by hand drafted features.
Where you are in the process now, how soon do you think this will be available for actual public use for doctors to use?
BAGCI: This is already available for research use. They are also talking with UCF commercialization office to know how to make this product and this will require FDA approval for clinical use but currently it is already available for research purposes. And what we are doing now is increasing the dataset and we are improving the algorithm to improve the accuracy sensitivity and specificity method to get it a bit higher. The next step is going to be FDA approval and it will go to the clinic but it may take some time. However it is available for research use now already. Our source code (algorithms) are publicly available as well and we are also testing our records in the publicly available tomography scan so everybody can also get those scans and test our algorithm and they can see how it works.
How can people test it?
BAGCI: Yes, they can download our codes because they are publicly available and they can use also publicly available images, tomography images, they are available. Even high school students are actually getting this data and testing some lung cancer detection problems. We are using them for education purpose as well here.
And what other projects are stemming out of this system? I know you’re mentioning that you’re working on pancreatic tumors.
BAGCI: Yes. We have an ongoing successful collaboration with Mayo Clinic in Jacksonville where we are developing a similar system for pancreatic tumor detection. And then diagnosis. Our idea is to find the pancreatic cyst before they turn into cancer because this is the only way as of now that we can prevent pancreatic cancer. So when pancreatic cancer is detected, usually it is very late stage. That’s why we are focusing on pancreatic cysts (pre-cancerous tumors). If we find the cysts automatically and if we can decide if cyst has a high potential of turning into cancer, this may be a really groundbreaking and we already had very good results. I think our group is the first one in the world using MRI and AI to diagnose risk of pancreatic cancer. We already filed our patent and we have a couple of papers. We will also release the results.
And is that is that currently used for research purposes now or is also in the process of being FDA approved?
BAGCI: It’s only the research purpose for now. But the thing is, there is no single system doing this yet. This is the first ever system. It needs to go to the FDA for approval and then it’s going to be approved for clinical use afterwards. But here’s one important fact; there is no system available for pancreatic cancer screening for you in the way that we do that. There are several companies, several researchers are working. But the accuracy was so inferior. If you go to the hospital or research center you will see that most of these artificial intelligence solutions are not adopted widely but our system is putting the radiologist in the center. So what happens is that our results can be explainable and we are going beyond the black box nature of the artificial intelligence so radiologists can trust our tools. In many times they don’t trust computational tools because it is spitting just numbers saying that this is cancer or this is not but our system goes beyond this.
Are there any limitations to using this system?
BAGCI: Yes, the accuracy is not 100 percent but this is within the radiologists’ interpretation level, even better than international standards. If we have larger amount of imaging data we believe that we can even surpass the current accuracy. And so this is a developing field still and we are close to 80%. We will continue to improve that. That is the limitation.
So basically the more data the system has, the better.
BAGCI: Yes if the data is well-curated. We believe that the variation is the key here. Having very different scans from very different types of people such as gender race and different parts of the world. We are now contacting researchers in China to get some tomography images to increase the variation in our data. So that will improve generalization properties of our algorithms.
What indications do you think that that would have in the medical world?
BAGCI: The radiologist will read the scans for the first time and will capture those missing tumors. It is going to improve the life span of people because you are capturing the lung cancer or pancreatic cancer at the early phase. For example, for pancreatic tumors if you find it. And our algorithm tells you this is having a highly likelihood of turning into cancer which means the surgery is the option immediately. The surgery can be done and then you will be tumor free. Similarly early detection of lung cancer; if it is later stages our algorithm can tell the prediction of the survival rate and prediction of how this tumor is going to grow. The algorithm is helping generate better treatment strategies for the later stages such as the surgery modelling, where to cut, calculating the radiation, those calculations, all those things. So it can be used for many purposes starting from surgery to different therapy strategies like surgery and radiation therapy and prediction of survival.
So it doesn’t just spit out whether it’s a cancerous tumor or not. It also tells you the stage?
BAGCI: Yes. We have the accompanying algorithm when you have the PET images in the later stages of the lung cancer. This is the positron emission tomography. So we are able to actually tell stage of the tumors, how much dosage should be delivered to the patient, and what kind of therapy should be done based on the stage because we are able to understand them based on the walling of the tumor and where it is. We are guiding radiation and radiation therapists to a personalized therapy and to maximize the benefit, killing all the tumor tissues while minimizing the killing of healthy tissues. So it is really helpful for optimizing the therapy plan.
Anything else that you think people viewers should know about this system?
BAGCI: Yes. We are very excited about this system. We are looking for more partners to have more data. We can collaborate with them to test our system in the hospitals and research centers. And then this is going to facilitate the adoption of this tool in different centers. We want to really focus on the research side and there are so many things going on there. But the system is ready to be tested.
Do you think it’s maybe like a year, two years?
BAGCI: Yeah maybe in the past it was taking longer but nowadays we are seeing that these computational tools are getting less than a year for approval.
END OF INTERVIEW
This information is intended for additional research purposes only. It is not to be used as a prescription or advice from Ivanhoe Broadcast News, Inc. or any medical professional interviewed. Ivanhoe Broadcast News, Inc. assumes no responsibility for the depth or accuracy of physician statements. Procedures or medicines apply to different people and medical factors; always consult your physician on medical matters.
If you would like more information, please contact:
Robert Wells, Media Relations
407-823-0861
Sign up for a free weekly e-mail on Medical Breakthroughs called First to Know by clicking here.
