Canon Medical Research Europe has been given £140,000 to develop a prototype that combines artificial intelligence and medical imaging to improve the assessment of one of the most difficult cancers. The firm was awarded the Phase II funding by the Cancer Innovation Challenge and is working with NHS Greater Glasgow and Clyde to improve the assessment of Malignant Pleural Mesothelioma (MPM) tumours, also known as "asbestos cancer".
DataIQ spoke to Dr Kevin Blyth of NHS Greater Glasgow and Clyde and Dr Keith Goatman, principal scientist at Canon Medical Research, who explained that MPM tumours are tricky to identify because unlike other cancers, these tumours grow around the lung.
"We are developing an AI tool that can find a tumour on scans."
This makes is difficult for radiologists to measure the tumours on scans. Blyth said: “We are trying to develop an automated AI tool that is capable of finding the tumour on the scan accurately, then measure the volumes of that complex morphology or shape, and then compare it between multiple time points.”
He said that by doing this, they would be able to see if a particular cancer drug has been effective. The doctors are hoping that the AI assessments will be more robust and less variable than the measurements made by human radiologists which can vary by up to 50%.
Phase I of the project was a three-month period of demonstrating that it was feasible. Blyth said: “We set ourselves some goals. There’s quite a lot of a governance to navigate here in terms of access. All the data needs to be anonymised. There needs to be robust ethics and governance, so all those pipelines needed to be established through Phase I.”
"We plan to compare the performance of the AI to the human."
Blyth pointed out that he is running a concurrent study called Prism which involves looking for a gene signature that predicts whether patients respond to chemotherapy. He said: “In that study is where we’re getting a lot of images. Every patient in that study is having their CT images reported up to three times by different radiologists. So we’ve got a robust, gold-standard result. What we plan to do in this project is compare the performance of the AI to the performance of the robust human result.”
Goatman went into more detail of the three tasks that took place in Phase I. The first was to anonymise the data of the scans that the AI was trained on. The second was to investigate possible solutions and potential techniques that they could apply once they had the data. The third thing was to annotate the data by getting humans to indicate on the images where the tumour is and that information was given to the computer to help it learn to detect tumours by itself.
“At the end of Phase I, we didn’t have a working prototype. There wasn’t time to do that and that’s what we’re intending to achieve in Phase II,” said the principal scientist. The second phase is expected to take six months.
The scientist stated that the most challenging task was the technical one of detecting the tumours. “A tumour as irregular as this one is not spherical, so you cannot measure the diameter, and you have to detect every single bit of it to determine how big it is.” He added that not only the shape, but the contrast make the tumour difficult to detect as it doesn’t stand out and it is very hard for a human radiologist to differentiate tumour anatomy from non-tumour anatomy.
"This is the most significant jump in image analysis performance in 25 years."
Taking a deeper dive into the technology that is powering this study, Goatman explained that: “AI is the overarching technology and part of what’s doing that is image analysis. The key technology we’ll be using is deep learning, a type of neural network.”
For him, working with this type of technology with this level of computing power is one of the highlights of his career. “I’ve been working in image analysis for 25 years and this is the most significant jump in performance I’ve seen in those years. This one technique just blows what we’ve done before out of the water, but it does depend on having a lot of data.”
In the past, he and his colleagues would make models of how an algorithm would work and apply their thoughts to that process so they didn't need as much data. However, now that they are modelling their algorithms on real data, “it just beats human developed models hands down, time and time again.”
This use of advanced technology to identify MPM tumours is necessary because the UK has the highest rates of mortality in the world. Blyth said this was down to prior use of asbestos in industries such as ship building. Cancer Innovation Challenge is an initiative aimed at stimulating novel data and technology techniques to help Scotland become a leader in cancer care. Canon Medical Research is one of three initial recipients of funding in March.