

Hello, and welcome to UroToday’s Bladder Cancer Center of Excellence. Professor Kamat is at the Anderson Cancer Center. It’s a pleasure to welcome Dr. Hikmat Al-Ahmadie from Memorial Sloan Kettering, who is a Chief Pathologist and Associate attending in the Department of Pathology, and Professor Olivier Elemento, the Director of the Englander Institute for Precision Medicine. They presented data on artificial intelligence in cancer with specific reference, in some ways to bladder cancer, at the recent think tank, which concluded about a month ago. I’m very happy that both of you are able to share your wisdom and insight with the audience. I will give the stage over to you guys.
Thank you so much, I appreciate it. It’s a pleasure to be here today and I’m looking forward to having a discussion with you. I think the format of the discussion will need to involve Dr. Al-Ahmadie and me. Some exciting applications of artificial intelligence for cancer and bladder cancer were presented in a few slides that we prepared. There is a lot of excitement around the use of artificial intelligence in medicine. To give examples of applications, have discussions about the strength and limitations, and hopefully address some of the questions that have been raised multiple times, is what we want to do.
We’ll begin with a few slides and applications, and then we’ll discuss. These slides are meant to illustrate some papers that have come out in the field recently, that are really exciting, and they show how we are able to generate a tremendous amount of information as you know about each individual patient. We can use surgical samples to do this. The ability to sequence almost the entire genome of tumors using technology such as high frequent sequencing is one of the exciting things that has been happening in the last few years. We now have the ability to sequence the genome from beginning to end and get a lot of insight into individual patients. More and more this information is being used to treat patients.
The challenge is that when we sequence a patient’s tumor genome, we find a lot of mutations, and the challenge is to interpret those to guide treatment and diagnosis, based on this very complex information that we see in every patient. This is an example of how we can use the data. There is a recent paper that shows how we can use information from patients. This is data from many patients. We’re showing how we can use this information to predict the type of tumors. This shows a connection between the disease and the tumors. If you combine several genes that you see in patients, you can tell if a patient is a breast or bladder cancer patient.
A combination of genes adds a lot of value in indicating the type of cancer. There’s a lot of interest in using this data as a way to predict the types of tumors for patients whose primary type of tissue is not always clear. The amount of precision that you can get from the profile of patients is shown in the slide. If a patient has lung cancer. Very precise, you don’t make many mistakes when you can quantify the content of hundreds of genes. One of the applications is prediction of the tumor type for cancers, but non-primaries, for example. This is great because it is an illustration, but it is also great because it shows the power of thegenomics.
We’ll discuss this in the discussion. It is a very broad interest in using genomic data to predict response to treatment. This is the beginning but there’s a lot of work to be done. Predicting response to immune therapy is an example. It takes more than one gene to predict an immune therapy response. It could be a combination of genes or a combination of genes. Predicting response or diagnosis using data from multiple genes, multiple modalities, transcriptomes, and genomics is the future. This is just an illustration, but we will get to this more in the discussion.
Hikmat Al-Ahmadie, thank you for this opportunity to partner with Dr. Elemento, and to show you some of the interesting and exciting things that are happening in Artificial Intelligence. There are huge applications for artificial intelligence and deep learning in the diagnostic realm of care and cancer care, which is why I shift my focus to the diagnostics of bladder cancer. The examples shown by Dr. Elemento show you that you really need an important requirement for implementing artificial intelligence and for its success. If you don’t have that type of infrastructure, you can’t have artificial intelligence. It takes a while until we are able to get a complete image of a slide, so that’s why it’s so natural to pathology applications. It is being used in medical centers, small and large, throughout the world, as it is becoming more and more mainstream. There is a great opportunity for any type of artificial intelligence.
The ability to provide diagnostic accuracy and reproducibility will result in high-level efficiency in assessing these pathological materials. It will be very important in the future, or even happening now, to incorporate any potential application for biomarkers, whether that is for diagnostic purposes, for prediction, or prognosis, all of that, are a ripe environment forAI to be implemented.
You can have a high level of accuracy in predicting a site of origin, if you use the example that Dr. Elemento mentioned. The authors of this seminal paper tried to do the same thing by using a deep learning and artificial intelligence approach to determine the site of origin, but they were able to train and verify their results by looking at the slide images. One area where bladder was one of the tumors that were assessed achieved a good level of accuracy.
In one study, the authors tried to reproduce the grade of urothelial carcinoma by comparing the automated grade generation with that of three pathologists who themselves had a moderate agreement. They were able to reach a moderate agreement with the consensus grade given by the pathologists. The authors were able to achieve a very high level of accuracy by using a deep learning approach to distinguish between the two diseases, and they were able to reach 98% of the cases.
In another study, the authors were able to come up with a method that could expedite the identification of bladder layers and potentially help in the diagnosis of T1 disease. By applying deep learning approach to urine cytology, the author in one study was able to use digital image analysis to find better samples than a simple histopathology review.
Artificial intelligence has the ability to predict a genetic abnormality starting from the histopathologic slides, as has been shown in some publications, by a strong correlation with and detecting FGFR3 mutations, and another study by the ability to detect the different types of bladder cancer. Some of these applications have the potential to be applied clinically and enough in the distant future as shown in this study, where the authors were able to show a good level of accuracy in making a diagnosis of cancer that matched the pathologic evaluation,
There’s a huge potential for artificial intelligence and deep learning in medicine, and specifically in bladder cancer. Despite the exciting results that have been reported so far, a lot of them are still proof of principle and there is still a lot of validation that is needed. If you have a rare disease with not always a straightforward presentation or features, there are potential limitations to an artificial intelligence approach. Bladder cancer is a good example and many of its states may not be straightforward but I think these things can be overcome. Since it’s a trainable approach, it’s important to have accurate algorithms that will provide the input for the training sets before it can be applied for validation. We’re happy to discuss with you, thank you.
Ashish Kamat is happy. Thank you both for the great summary of the talk. The conclusion slide covers a lot of the points I wanted to raise in my discussion. I’m going to raise that, so we can discuss it. I wanted to ask you and both of you if there’s a perception among people that aren’t involved in the field that we could just now have artificial intelligence take over. What are some of the practical hurdles that need to be overcome in order for artificial intelligence to be able to replace the pathologist in day-to-day practice?
This is a real fear and it comes up all the time. It’s not necessarily a fear, but it is a worry that what is the role of the artificial intelligence and is it going to replace pathology? As I mentioned in my slide, having image recognition is very important and computers can perform very well, and pathology, as I mentioned in my slide, can lend itself naturally to digitization. It’s better to embrace it and see how it can improve what we do, instead of seeing it as a competition or threat.
There is no doubt that skilled eyes are very important in identifying diseases. If it is trained well, it can identify a lot of the routine, straightforward cases that may not need the same level of attention as other cases that might require more human input. I think that using artificial intelligence in the diagnostic world can save a lot of time and effort. No matter how you look at it, it is more efficient and more reliable than humans. I’d rather use that to our advantage, let the artificial intelligence help and expedite the routine diagnostic process and leave the more challenging cases or scenarios that would require human intelligence to be handled by the pathologist.
One place where I think there’s a real role for it is in the unknown primary clinics that we have, right? Patients with no known primary source are the ones Dr. Elemento shows in hisJAMA Oncology Publication of the Tumor Type Prediction. Is it possible that a large benefit of an artificial intelligence, gene-derived tumor type prediction would be in a specific group of patients?
Hikmat Al-Ahmadie: Absolutely. That would be very helpful. We saw that from the digital slides, but I think Dr. Elemento has more to say about other approaches.
And that’s alright. I wanted to add a couple of comments to what Dr. Al-Ahmadie had said about the use of artificial intelligence in order to help experts who are often very busy. I think it has the potential to help some tasks. I think it’s important to remember. I believe we are in the early days of using artificial intelligence for medical applications. Right now, artificial intelligence can’t really explain itself, which is a major shortcoming of the technology. The inability of most artificial intelligence systems to explain why they are making a particular decision is a major problem. There is a critical gap in medicine.
The Pathologist, like Dr. Al-Ahmadie, doesn’t just say, “Well, this is the diagnosis.” They can explain why they made the diagnosis. To be able to explain why the decision was made in the first place is absolutely critical in the conversation between physicians and patients. This is something that the field needs to work on. There is potential for improvement, that’s what I think. I think that at some point it will be able to explain itself. This is a reason why artificial intelligence isn’t always being used in the clinic or in pathology as much as it could be. It’s the inability to provide interpretation that’s limiting.
You said you were talking about the second point, so I’m going back to that point. Digital pathology is something I like because I think there are millions of slides that are being uploaded. There is so much potential to learn from a training slide because it is so accessible. We need to work a bit harder now that we have the data. There’s some work that needs to be done. There’s so much value in the information when we can get it and it’s easy to get it. The ability to get a precise diagnosis is one of the reasons there’s value in the diagnostic area.
There’s value in potentially matching a patient to therapy. Many of these are connected to specific therapies. That’s valuable information in bladder cancer. The presence of an FGFR2mutation means that patients are eligible for FGFR drugs. This is the beginning of being able to match patients to therapies that are based on their genes. There is value in the diagnosis abilities but also in matching patients to therapies.
I think the inability of artificial intelligence to explain itself is a negative, but it’s not necessarily a negative. If you want to call it pattern recognition, it can recognize certain patterns that we don’t understand. The tumor biology that we would learn in the future may be something to it. Is it possible that artificial intelligence will throw out some clues for people to start doing mechanistic investigations?
That’s an excellent point, I think. I think this is a point that we don’t hear a lot, as we learn how to look under the hood, if we want to be an artificial intelligence. I think we’re starting to see that artificial intelligence is able to transform stuck information into diagnostic information. That is really important. When an artificial intelligence is making a diagnosis, it’s looking at certain patterns and cells. We need to learn how to understand what it’s looking at and how to give it valuable information.
If you can look at the weights of sudden connections in a multi-layer, deep learning algorithm, you can see where the machine is looking at. You can get a map of what we call an attention map, which is basically a heat map of the people that you are looking at. I think there is a lot of useful information.
It’s difficult to transform this information into something that’s semantic. Artificial intelligence tells you that a bunch of pixels is important, but that doesn’t always translate into something that we can understand, like a type of information that we can use. There is work being done to connect those two worlds, but these are the early days, I think.
This is an exciting topic that we could talk about forever, but in the interest of time, we need to wrap it up. Let me give the stage back to you in some way, and then Hikmat, and then Olivier, have you kind of leave our audience with some high-level closing thoughts that you want to share with them?
Hikmat Al-Ahmadie is correct. It’s clear that artificial intelligence is already transforming how we think and how we do things, and it’s going to be a tremendous help on many levels, but it still has some challenges to overcome. Even at the current stage, it will still require a lot of input from us, especially with the unusual and rare tumor types. There’s a huge potential for it and it should be embraced on all levels.
I think that seeing what’s happening in the field of bladder cancer is something that I have. There are many novel therapies being approved. There are so many therapies that are close to being available that I think these are exciting times in bladder cancer. I think the next step is how to combine those therapies, how to match patients to therapies, once you have a lot of reach in terms of therapies, you really need to think hard about who’s going to benefit the most from different therapies.
I think that Artificial Intelligence is going to help. I think being able to predict who will benefit from a therapy based on combinations of markers is going to be the future of artificial intelligence. The integration of multiple signals and being able to use this integration as a way to make predictions that are reliable and robust is one thing that artificial intelligence is really good at. I think these are areas where artificial intelligence is likely to be making an impact in the future, and I think that is exciting.
Excellent, Ashish Kamat: Great. Thanks again, both of you, for taking time from your busy schedule to share your thoughts on this important topic. We’ll get a chance to see each other soon, because it makes a lot of sense, and I’m getting tired of the zooming, but stay safe and stay well.
Thank you, Hikmat Al-Ahmadie.
Thank you for having us.