Main content

How AI could transform colorectal patient care by creating personalized treatment pathways

Gary Flood Profile picture for user gflood April 15, 2024
Summary:
Using IBM AI-as a-service from managed service provider CSI, Oxford University spin-off Oxford Cancer Biomarkers thinks it can help identify patients who really don’t need the challenge of post-surgery chemo

An image of a scientist in a lab analyzing images on a computer
(Image sourced via Oxford Cancer Biomarkers)

IBM Artificial Intelligence (AI) technology is helping a UK precision cancer diagnostics provider help deliver better clinical decision-making and improve patient outcomes.

The company - Oxford Cancer Biomarkers - is using what it calls a new “digital pathology” approach to create a unique, personalized risk profile for patients.

By detecting cancers that are most likely to reoccur, AI is helping identify the 8 per cent of patients who will benefit from chemotherapy, but also help identify those who could possibly avoid tough follow-on post-surgery treatment.

The application of AI also means that treatments that better suit patients suffering from bowel cancer can be more accurately and rapidly identified, says the company’s Chief Operating Officer, Guy Mozolowski.

But by using AI computation power delivered by its MSP partner, the team can now think of a problem on day one, organize and curate its data on day two, run a model on day three, and then validate results on day four.

That means work that would previously have taken three to six months can now be prototyped in a week.

R&D development speed also means that, as a growing UK biotech, the company can ‘fail early, but fail cheap’ - jettisoning ideas that don’t have long-term potential and then put resources into more promising ideas.

That’s important - as ultimately, we’re talking about what could be life or death decisions here.

Mozolowski says:

The working definition of personalized medicine is the right drug or the right treatment, for the right patient, in the right dose. What we’re aiming to do is provide the diagnostic tools that will allow clinicians to stratify their patients into groups that they can then tailor treatment to, rather than treating them all as one group.

Improving long-term post-cancer surgery survival rates

At the heart of turning the dream of personalized cancer treatment is the computerized analysis - at scale - of images of a small piece of biology, called a biomarker.

A biomarker is a change that can be used to identify and monitor pathogenic processes or responses to a pharmacological intervention.

In cancer diagnosis, biomarkers are typically used to find changes in tissue samples that can point to more targeted and personalized treatment.

Spun out from the University of Oxford in 2012, Oxford Cancer Biomarkers has chosen to try and apply the use of biomarkers - initially, to help fight colorectal cancer, though it plans to extend that target set.

The main Oxford Cancer Biomarkers product that makes use of AI to do this is a software tool called OncoProg. 

OncoProg uses intense scanning of imagery to help clinicians decide if a patient would be better treated by surgery - or, in more high-risk patients, if they would be better off with what doctors call “adjuvant” (additional, post-initial intervention) chemotherapy.

He says:

Currently, the majority of mid to late-stage bowel cancer patients (Stage IIb disease and higher) are offered chemotherapy that is given after surgery and has been shown to improve 5-year survival rates of bowel cancer patients by up to 8% compared to surgery alone.

There are, however, disadvantages to this approach, as clinicians can end up giving a lot of people chemotherapy to help a relatively small number. So, a lot of people are given chemotherapy who do not benefit from it but are still at risk of significant side effects.

We thought, wouldn't it be nice if we could find and identify that 4 to 8% of people a priori, only treat them with chemotherapy and spare everybody else who's unlikely to benefit from chemotherapy from rather a torrid time and an unnecessary application of some quite powerful drugs.

AI models trained on images 

Supplying the AI firepower to do all this is a Birmingham-based managed service provider called CSI.

For its work with Oxford Cancer Biomarkers, the company is hosting AI as a service on a server provided off its Microsoft Azure cloud. 

This service is based on the IBM Power System AC922, combining 40 Power CPUs, with four NVIDIA Tesla V100 GPUs, running Red Hat Linux. 

The Power AC922 server is supposed to be specifically designed for deep learning, AI, high-performance analytics, and high-performance computing (HPC).

In this use case, NVLINK2 technology provides direct CPU-GPU communications for high performance, while an IBM V7000 is delivering 14TB of SAN-attached storage for the datasets, and AI models Mozolowski’s teams want to build.

Large Model Support is also being offered to the company - allowing its AI models to be trained on images larger than the memory available on the GPU, by using system memory.

Building and perfecting the deep learning model training tools, like OncoProg rely on, is also made easier, says the partner, by a distributed training feature that allows jobs to be broken down and processed in parallel over a cluster of GPUs. 

However, AI was not in the original Oxford Cancer Biomarker plan, says Mozolowski. 

He explains:

It would be wrong to say we had this all planned out years and years ahead; it’s all developed very organically, with us progressing from being a small, very academically focused organization, into something that's far more commercially oriented.

Back in 2021, Mozolowski says his developers started to see limits in what they could do, in terms of processing the images they needed to work through rapidly, in order to make the difference the company wanted to offer clinicians.

He says:

We had seen an opportunity to bring the power of AI together with image analysis and our domain knowledge of cancer biology, pathology, and oncology, but we very rapidly hit a wall in terms of what we were capable of in-house.

Each individual patient sample can be up to three gigabytes of data, with very high resolution electron microscope images.

As Oxford Cancer Biomarkers was dealing with image datasets of 2,000 to 3,000 patients, this quickly meant being able to handle hundreds of terabytes.

Mozolowski adds: 

We realized that to do things at the speed and the scale that we desired, we were going to have to partner with somebody who could provide a much, much larger scale of infrastructure, data storage, compute power and whatnot.

There was also market pressure to find an answer, says Mozolowski. He adds:

We saw that if we were to try and do this work with on-prem hardware, or what we had in hand at the moment, it would take us six months to run a basic PoC project - but with the speed that the market was developing at, we needed to be there in more like weeks.

Potential NHS cost savings

After market analysis, that technology emerged as IBM’s, delivered by CSI.

He says:

What really sold us with these guys was just the level of service, the level of expertise, and the can-do attitude.

They're also very nice people, which always helps.

Next steps for AI and biomarker-based digital pathology at Oxford Cancer Biomarkers will be close monitoring of the system’s use at two NHS Trusts.

Mozolowski says:

We know the technology works, but we are essentially running what we call a ‘clinical utilities’ study to prove it delivers all the benefits that we have seen in our very controlled trials in a real-world environment, with all the vagaries of real-world clinics and patients.

We’re looking closely at performance, the outcomes for the patients, and potentially all the cost savings to the NHS and other health systems of using AI and biomarkers like this.

Loading
A grey colored placeholder image