UK govt plans cancer AI, Intel turns to ML-powered pharma

The UK is planning to leverage its NHS to improve its cancer diagnosis procedures, by combining its vast array of data with AI-based technologies to spot early signs of cancer. Combining medical records with population data, the hope is to prevent 20,000 cancer-related deaths annually, by 2023.

It’s ambitious, but the UK is well-placed in such projects. Critics will argue that the current government’s economic austerity policies and its budget cuts to the regional NHS trusts are a bigger hindrance than a lack of AI, and that it should look to rectify those aspects before trying something shiny and new.

Nevertheless, prime minister Theresa May has called for the NHS, the emerging AI sector, and health charities to collaborate – pooling data to spot the disease’s early warning signs. The government believes that new approaches could lead to 50,000 annual early diagnoses of bowel, lung, and prostate cancer – which would lead to many of those potential cases being treated before they are fatal.

The funding will come as part of the £1.4bn Grand Challenges program, which aims to tackle four emerging global trends – AI and the data economy, clean growth, healthy ageing, and the future of mobility. With healthy ageing, the overall goal is to ensure that citizens enjoy at least 5 extra years of healthy independent life by 2035. Some £98m has already been invested, but there is £210m available for investment in the Data to Early Diagnostics and Precision Medicine pot.

One of the major cancer charities in the UK, Cancer Research, says there were 166,000 cancer deaths in 2016 – around one in four deaths. The organization believes that if AI-based systems could reduce late diagnosis by half in the next 15 years, for four types of cancer (bowel, lung, ovarian, prostate), then it would avoid 22,000 annual deaths.

But of course, cancer isn’t the only condition that could be tackled by new analytics capabilities – nor is it a singular condition, as there are many varieties. Diabetes is one of the most burdensome conditions for the NHS, due to its lifetime of complications, and an aging population is increasingly encountering degenerative brain diseases, like Alzheimer’s and dementia.

For these AI-powered analytics, this means that there will likely be a different approach for each type of cancer – one for lung, one for bowel, and so on. This also means that patient data would have to be parsed through each system in order to ‘test’ for each different type. Depending on resources, this could be quite inefficient, and it seems likely that other filtering systems will need to be used to allocate patients to these tests.

Medical images, like X-Rays and MRI scans, are another area where there is great hope for AI-based systems. A recent study has just outlined how a CNN outperformed dermatologists at identifying cancerous moles and benign spots – achieving a 95% accuracy compared to the humans’ 86.6%.

In her speech, the PM said that these technologies “open up a whole new field of medical research and give us a new weapon in our armory in the fight against disease.” She added that “achieving this mission will not only save thousands of lives. It will incubate a whole new industry around AI-in-healthcare, creating high-skilled science jobs across the country, drawing on existing centers of excellence in places like Edinburgh, Oxford and Leeds – and helping to grow new ones.”

Elsewhere in the medical sector, machine-learning has shown great promise in the R&D of new drugs, with pharmaceutical companies keen to use such systems to do a lot of the pre-trial legwork. Simulating how molecules behave and interact with the human body can allow researchers to find new potential combinations and approaches.

To this end, Intel has been working with Novartis, one of the largest pharma companies, to explore how a DNN system can improve ‘high content screening’ – a process that analyzes how potential candidate drugs interact with human cells, using a combination of tests and screens to determine how a drug works.

The pair say that their collaboration led to a 20x improvement in the training time needed to get the DNN up to speed, using image analysis models. The tests used 8 CPU-based servers running an optimized TensorFlow session, to process the 10K resolution image dataset – the Broad Bioimage Benchmark Collection 021, to be precise.

The ‘high content’ part of the screening refers to the features needed to be analyzed, which include the size, shape, and texture of the cells. With the HCS system, thousands of effects can be studied more quickly. This is where the ML function comes in, with Intel and Novartis saying that the DNN has the potential to automatically distinguish the effect of different treatments on the cells.

The very detailed microscope images are much larger than those used in the ImageNet datasets used to train image recognition systems. In turn, there are millions of parameters in the DNN model to account for, and thousands of test images. Intel says this adds up to a high computational load, which then means that such training takes a long time.

But this is where the DNN Acceleration comes in, which Intel and Novartis say has led to that 21.7x improvement – from around 11 hours to just 31 minutes. However, it should be noted that the claim is based on scaling the system from a single Xeon Phi 7290F to that 8 CPU cluster – specifically, 8 Xeon 6148 processors (20 cores each, supporting 40 processing threads each).

More details of the hardware configurations can be found at the bottom of this page, but it’s important to note that the cluster had 8 times the RAM of the Phi, which is largely why it could process the images more quickly. The pair conclude that the new approach provides a quicker result, and one that captures more information from the images – apparently able to process more than 120 3.9MP images each second, using 32 TensorFlow workers.