September 13, 2018

#S02E02 - Can AI be used to detect cancer?

Companies are turning their artificial intelligence networks towards cancer detection and treatment. So can AI really be used to detect cancer?

#S02E02 - Can AI be used to detect cancer?

This year 1.7 million cases of cancer will be diagnosed in the United States, and for around 600,000 people, that will be a death sentence. But as technology improves companies are turning their artificial intelligence networks towards cancer detection and treatment. So can AI really be used to detect cancer?


This episode of Moonshot was hosted by Kristofor Lawson (@kristoforlawson) and and Andrew Moon (@moonytweets).

Our theme music is by Breakmaster Cylinder.

And our cover artwork is by Andrew Millist.


Woman: “My mum was on the phone with the doctor, she said ‘it can’t be cancer’.”

Woman: “My doctor said, you had cancer.”

Newsreader 1: “According to the American Cancer Society, one in seven men will get prostate cancer.”

Newsreader 2: “Singer-songwriter Olivia Newton-John revealing her latest health battle, fighting cancer for a third time.

ANDREW: Cancer… it’s possibly the worst thing you could hear when you go to the doctor. This year - one point seven million cases of cancer will be diagnosed in the United States… and for around six hundred thousand, that number that will be a death sentence.

[News Report - “IBM Watson may be the technology that helps doctors beat cancer - ]

KRIS: But new technology like IBM’s Watson computer is helping doctors detect and treat cancer at an earlier stage…

Watson Promo: Watson for oncology can help identify and reduce variability by providing your clinicians with information regarding evidence based patient-centric treatment options.

KRIS: Welcome to Moonshot - the show exploring the worlds biggest ideas and the people making them happen. I’m Kristofor Lawson.

ANDREW: And I’m Andrew Moon.

KRIS: And in this episode of Moonshot we’re exploring some of the innovative solutions that doctors are using to help detect cancer, and the impact that AI could have on the outcomes for cancer patients.

ANDREW: But before we dive into the details, here’s a word from our sponsors:

Dishan Herath:  So there's four stages for most cancers.

KRIS: This is Dish Herath… he’s a medical oncologist at the Peter MacCallum Cancer Centre and also Western Health in Melbourne, Australia. And he’s on the front lines dealing with cancer every day, and a lot of the patients that he sees suffer from lung cancer.

Dishan Herath: For lung cancer, a lot of patients will present in stage four which is the most advanced stage, which generally means for a lot of those patients, there's not a cure available, but a percentage of those patients can actually be cured now. So not a large percentage, but some can be cured. Certainly for the earlier stages, there's more chance of cure. So stage one and two, there's more chance of cure, stage three is kind of in the middle. So that's the sort of situation we're dealing with. But unfortunately, we do get a lot of stage four lung cancers.

KRIS: Right now, cancer detection relies on doctors or patients noticing lumps or spots on their body, or they might be found via traditional MRI or CT scans. But unless you’ve got a reason to believe something is wrong - often you may not even be on the lookout for symptoms that could indicate a cancer diagnosis.

Dishan Herath: For an external cancer like a skin cancer, so if you've got skin cancer or melanoma, it very much is about picking up that there is a suspicious spot, essentially, or a lesion on the skin, and the key factors around that, that it's changing over time. So it's first of all that it's got some suspicious features that make it more likely to be malignant, well, cancerous rather than benign, and secondly, that it's changing over time... Now, if you're thinking about something like an internal cancer, it's very much dependent on where the cancer is in the body and the sort of symptoms it will present with. So for prostate cancer, that'd be that it might be affecting how men are able to pass urine, how frequently they're going to the toilet. For something like lung cancer it'll be, are they coughing more, are they getting more chest infections, those sorts of things.

Dishan Herath: Now, there are screening procedures first. So if you're talking about prostate cancer, there's the PSA blood test, though that is not perfect. And for lung cancer, there is some work being done looking at screening smokers with CT scans. And again, that's still an area of active research and some controversy, I think.

KRIS: Medical scans like CT and MRI have made cancer detection far more effective in the past 50 years. They have revolutionised treatment of cancer and improved patient outcomes - even for the most severe cases. But now there are new ways of using this imaging data to help patients.

ANDREW: Previously on Moonshot we’ve spoken about this rise of artificial intelligence in areas like robotics. But it turns out that AI and machine learning techniques can also be applied to healthcare. And specifically, cancer detection makes use of a lot of medical imaging - so there’s potential to use all of that data to train an AI system to recognize patterns, and help improve the cancer detection process.

MetaOptima Video: “Do you have moles? Moles start harmless but even if just one is cancerous i can be fatal.”

KRIS: MetaOptima is just one of the companies working at this intersection of AI and healthcare.

MetaOptima Video: “Here’s how it works, scan your moles using the molescope camera attachment, use our mobile app to securely share your scans with your doctor. Get an experts opinion from the comfort of your home.”

Maryam Sadeghi: So, the idea was how we can bring collective knowledge of thousands of doctors to have the better diagnostic and treatment decision when it comes to a new case. So, it was all about looking at the data, learning about similar cases and trying to understand and have a diagnosis of these skin disease or skin problems.

KRIS: That’s Maryam Sadeghi, a computer scientist who co-founded MetaOptima  with her husband in 2012.

Maryam Sadegh: Imagine, two computer scientists with training in medical and dermatology were excited about the future of data-driven, value-driven dermatology and we said let's make it happen. We can build it.

KRIS: MetaOptima is specifically focused on using AI to detect skin cancers and also melanomas, but rather than giving you an explicit diagnosis - the AI engine aim to give patients and doctors the knowledge that they need to classify what those legions actually are.

Maryam Sadeghi: The doctors told me, "Oh this is like my intelligent textbook." I was like "Awesome." This is the best, you know, simple brief that you proposition that I could actually think for DermEngine. Because, he was like, "Oh, I thought you might takes look for other similar cases." But this is like matching images with those, it's intelligent, it can show me, wow, 100 similar lesions and sometimes like, oh is it the same one, it's so similar. Because, this is simple for machine, it's like pattern matching, finding those cases, looking at the outcome for those cases as well.

KRIS: MetaOptima has two different branches of their scanning technology. The first is MoleScope and the second is DermEngine. MoleScope is targeted at patients - you can use it at home to monitor and track moles, and DermEngine is targeted at health professionals.

Maryam Sadeghi: DermEngine talks to any other device in the yard kit. You can use any device. Any imaging device, any big machine. We connect with those systems and give you centralised access to the data, to the patients that you have on the physician's side.

Maryam Sadeghi: On the patient's side, we actually decided to have affordable devices for them to be able to capture quality data. If we are talking about empowering patients with AI they need to have also the right tool to be able to give the quality medical data to the system, to the AI, that can help them.

KRIS: Essentially, the main aim for company is to incorporate artificial intelligence into the workflow of diagnosis rather than actually replacing the doctor. And Maryam says the reason for this is because AI systems, while they can be fast at doing tasks like image classification, when it comes to something as serious as cancer you don’t want any room for error, so you don’t want to put too much reliance into an AI system unless you can prove it works all of the time.

Maryam Sadeghi: This is not Airbnb, like yeah you can have disasters but at the end it's just money, it's fine. These are our patients lives, this is very different.

ANDREW: Now, AI isn’t limited to any particular type of cancer detection. So while MetaOptima is specifically looking at skin cancers,  which are very easy to see and take a photo of, other companies are taking on the task of identifying cancers that lie within.

Elliot Smith: So we looked around at a number of different cancers, and we found that prostate cancer has ... number one, it has a lot of medical imaging in it's pipeline. But that medical imaging has really only been taken up in a big way over the last sort of five-ish to 10 years. And we felt that there was an area here where we could provide a lot of benefit.

ANDREW: This is Elliot Smith - Elliot is the co-founder and CEO of Maxwell Plus - a company which is using AI to help detect prostate cancer in men.

Elliot Smith: Prostate cancer affects a lot of people. In Australia, in roughly one in every seven men will, at some point in their life, generally much later in their life, so 50 plus, will develop prostate cancer. So we felt that it was an area where our technology could be applied, but also where there was a need for some new technology to really help ensure that people were getting diagnosed early and treated in a low risk and effective way.

ANDREW: In a similar way to MetaOptima - Maxwell Plus is building tools to help give doctors more confidence in the decisions they’re making. The company has started with prostate cancer - but they’re also working on the detection of lung and breast cancers.

KRIS: As we’ve talked about on Moonshot previously, the quality of an AI system is based on having access to really good data... and when it comes to diagnosis of patients it’s even more important that the AI is fed good quality information. But how much data do you actually need to make the system work at a level where the doctor could feel confident in the decision that they were actually making?

Elliot Smith: As a general rule of thumb, you'd want about a thousand cases in each of the different categories you want to classify between in order to start to see some good results. So in our case, if we simplify the problem a little bit down to "yes cancer", "no cancer", we'd want about a thousand in each category. And then from there, you know, that's gonna give you early indications of the results, but it's probably not gonna be human levels of performance.

KRIS: Now Elliot says that as you try to improve the AI system and reduce the error rate it requires a lot more information. Every time you halve the amount of errors in the system and therefore improve your cancer detection accuracy, you need to provide 10 times the amount of data that you previously had.

Elliot Smith:  So let's say we started at 80% with a thousand in each class, and we wanted to get up to 90%, as a rule of thumb you can imagine that we'd want to get to about 10,000 in each class to reach that 90% mark.

Elliot Smith: And then 100,000 to get up to around 95. So you know, there's definitely a huge amount of data that we need here, and when we're talking about 3D scans of patients, their genomics, and their blood work, every single example that we train on can be in the sort of tens to hundreds of megabytes range that needs to be processed through this algorithm.

KRIS: And we’ll continue our look at the use of artificial intelligence in cancer detection, right after this break.

ANDREW: Welcome back to Moonshot - I’m Andrew Moon. And when we’re talking about cancer it’s important to know how cancer cells actually work in the human body.

ANDREW: In a healthy body, cells grow and multiply in a controlled way, following genetic instructions to make sure everything is in the right place. However Cancer is caused by a defect in these genes.

KRIS: This defect prompts an uncontrolled growth of cells. Some types of cancer stay in one location and grow slowly. However - Malignant cancers are aggressive and can quickly spread to other parts of the body, making them much harder to treat.

KRIS: The complicating factor is that cancers can actually present differently in different people from different populations - which Dr Herath says still leaves a big question mark over the effectiveness of an AI system when trained on any one population’s data.

Dishan Herath: Something like the U.S. and Australia would be much more comparable in that regard. I think where you're going to see some pretty big differences, places like Japan where they have a very ethnically homogenous group of patients who are quite different to Australian patients. So we're probably more comparable to somewhere like the U.S. than to Japan, for instance.

KRIS: So any systems that are detecting cancers would need to be trained differently for, like a Japanese patient compared to an Australian patient?

Dishan Herath: That's a great question, which we don't know the answer to yet. And again, this is why I would say the way they do this is to actually look at the performance of the systems in the real world. It may be that you don't, it may be that you don't need to train it on different populations, that it can get enough information from a U.S. set of patients that it can then more broadly apply it. But we shouldn't just take that at face value where we just drop in the U.S. system and start using it here. You actually have to test out whether that is working in a real world environment in Australia.

ANDREW: Detecting cancer early is one of the keys to a successful treatment, and it’s hoped that using an AI system will alert doctors to potential cancer faster, and help them start treatment earlier. Another benefit? These systems could also have a knock-on effect by helping doctors become more effective with the resources that they have.

Elliot Smith: There's a number of factors here, one is that if you look on a global scale, there is a shortage of doctors, worldwide.

Elliot Smith: And some countries have it a lot worse than others, but even if you look at parts of the US, there are certain states where they have around 50% of what their projected need for clinical expertise is. So there's certainly a need for more clinical brain power, but also as we try and detect these cancers earlier and earlier, the signs of the cancer being there become more and more subtle. And we need systems that are able to look across complex relationships between data that aren't apparent to the human eye, and reshape those in a way that helps clinicians understand that data, and can perform that earlier diagnosis. Because as they’re small and harder to detect we naturally need better and better tools to be able to find them in the first place.

ANDREW: The World Health Organisation estimates a shortage of 2.6 million doctors worldwide, mostly in developing countries. And with such a huge shortfall in the pipeline, it makes detecting cancer at an early, low risk stage in the traditional way, a pretty difficult job.

Elliot Smith: So our goal is to get there when it's very early signs of cancer. So for prostate cancer, it can be a slow growing disease in some people, in others it can be quite rapid. So sometimes it may take three to five years in order to develop into something serious. We want to catch it when it's at that first sign. Now that doesn't necessarily mean we're gonna go in and do surgery right away, but it we know it's there and we can monitor it, we can work with the clinicians to pick a time to intervene when it's gonna be most effective, in terms of stopping the cancer, but also pose the lowest risk.

ANDREW: Elliot hopes that with more information, patients can make better healthcare  decisions on their own. Instead of a single doctor planning your treatment, an AI system would work alongside healthcare professionals to shape the course of a patient’s long term health.

Elliot Smith: Yeah, we believe that in the future, patients will have much better control over their healthcare. There will definitely need to be more information given to patients to help them understand what this data means. We don't want this doctor Google scenario, where people look up their symptoms and cause themselves all sorts of worry, but we think that with the right access to data and the right access to patient education, we can really turn healthcare from what it is today, which is a very reactive system, you know, you get sick, you go in, you try to get better, to something that's much more proactive.

ANDREW: But despite the progress that is being made in trying to improve the accuracy of these systems - Dr Herath says it could take a while before AI becomes reliable enough to be considered useful in the fight against cancer.

Dishan Herath: So the first thing I'd say is, I think that this is actually very exciting because it has the potential in the future to help patients. So that's the number one thing that doctors would be focused on... I think that in terms of where the companies are at at the moment, I think this is still very much in the research arena. I'm aware that there are companies that are marketing their products for clinical use. My personal feeling is I don't think that for many of those, that enough research has been done to show that there's benefits to patients, which would make that worthwhile at this stage.

ANDREW: And one of the concerns that many have when they hear about AI moving into an industry - is what will happen to the jobs of those already working in it.

Dishan Herath: I think if you think about the broader medical community, there are some fears around this, around the kind of impact on their jobs and other change in their work roles, or in some cases, potentially replacing them. So I think my feeling is I don't think that doctors are going to be replaced any time soon by these sorts of technologies.

KRIS: But change is inevitable - we’ve seen that in many industries as technology advances. And when it comes to healthcare, medical imaging and technology has already revolutionised cancer detection and treatment, improving patient outcomes and saving lives.  Artificial intelligence could yet-again redefine healthcare as we know it, and while it may take some time to work out all the bugs and make it reliable, It’s up to the healthcare industry to embrace these new systems, not to fight against them.

Dishan Herath: I think that healthcare workers and working together with these technologies will be more powerful than the technology competing directly with health care workers. So that's the way I see it. Having said that, I think that it's highly likely that in many industries, changes will need to occur in terms of how you do your work, workflows. I think that that's extremely likely. So the first areas that would see that happening are things like radiology where people are looking at scans and x-rays. But to be fair, these are guys who've had many disruptions in their careers already. Things change very quickly, there's new technologies like MRI, even within the space in my career, things like MRI have come in. So I think people shouldn't underestimate the ability of the people who work in these fields to adapt to change.

Maryam Sadeghi: I believe is very, very important to health and implementing AI to our impact in our everyday life, is that to understand the limitation, and to understand the requirement, and to understand the safe implementation for this technology to help us, I don't think we should expect... Think of it as reasonable and it's not fair, really. And on the other side, we shouldn't fight this.

Elliot Smith: We set ourselves a goal, over the next five years, to try and bring artificial intelligence based analysis of medical data to a million people around the world.

KRIS: This is Elliot Smith again.

Elliot Smith: That's a big goal, but really that's where we hope to get. We want a technology that's accessible, that's affordable, and that's widespread enough that we can make sure that the best in class in analysis of medical data is made available to all these people, so that no matter where you are, you have the very best in diagnosis and very best in screening made available to you.