Dr. Rob Brisk, Clinical Research Fellow at Craigavon Area Hospital, Belfast, on what’s going to be hot in medical AI next year
Ah, the late December column. An upbeat recap of the year’s salient happenings. A jovial, here’s-what-you-missed of the weird and wonderful from the last 12 months. A snappy wrap-up with a note of festive optimism, preferably with a healthy dose of Slade lyrics.
No, you’re right – I don’t have the heart for it this year either.
So let’s move straight on to 2021, and what to look out for in medical AI:
Radiology. Computer vision has had the lion’s share of attention over the last couple of years, and that isn’t set to change any time soon. What’s new is a growing focus on frontline clinical deployment. Curated radiology AI “app stores” are emerging as a popular platform model that takes the pain out of procuring task-specific applications one by one. Existing big hitters like GE and Siemens are aiming to capitalize on their infrastructure foothold and global reputation to take pole position in the radiology AI market. However, they face competition from young companies who can pivot quickly in response to the ever-changing demands of such a fast moving field.
In the R&D domain, data governance continues to be a major obstacle for training new radiology models as it does for all other areas of medical AI. We heard some positive noises around “federated learning” in 2020, but lack of AI infrastructure within clinical data lakes prevented this from becoming a dominant global framework. Looking ahead to the new year, a number of leading clinical institutions are investing heavily in on-site AI hardware, and it might be that 2021 sees FL really take off.
Pathology. You would be forgiven for thinking that what works for radiology should also work for pathology. It’s all biomedical imaging, right? Wrong. Firstly – and perhaps surprisingly, for anyone not working in this field – widespread digitisation of pathology slides is a very new thing. Secondly, whole slide images are huge, to the point where they often need to be split into hundreds of tiles to fit into GPU memory. Reassembling the data to get the output you want needs some clever tricks. Happily, research in this field is coming on nicely and the value of AI-enabled analysis of pathology samples is becoming widely recognised. Definitely a big growth area for 2021.
Natural language processing. Ask 10 geeks what’s exciting in AI right now, nine of us will say NLP. The transformer-based GPT-3 from OpenAI really blew people away last year, and there’s reason to hope that 2021 will see biomedical language applications continue to play catch-up. To date, transformer-based models seem to improve linearly with the size of training corpora and the amount of available compute. So the ever-growing body of digital medical literature and dedicated medical AI research resources like NVIDIA’s Cambridge-1 supercomputer should fuel this fire nicely.
One of the holy grails of biomedical NLP is robust ‘relational extraction’ between named entities within medical literature. This is a key step in building queryable knowledge graphs (imagine being able to ask, “Alexa, which proteins are commonly overexpressed in coronary artery disease?” and getting a reliable response). Results are still a bit hit-and-miss at the moment but improving all the time. If you’re interested in this field, you might want to keep tabs on the BLURB leaderboard over the next 12 months.
Genomics & molecular simulation. Huge things are happening at the sub-microscopic level. Most people are aware of the speed with which the coronarvirus genome was sequenced, and the importance of this for expedient vaccine development. But you might have missed the story about DeepMind’s AlphaFold algorithm and its recent performance in the CASP competition. If so, make sure you catch up. It might be the most important thing you read this decade.
Potentially, the ability to predict protein folding based on amino acid sequences (the task for which AlphaFold was designed) was the missing piece of the most important puzzle in modern medicine. Technologies now exist that allow us to sequence DNA and RNA at a single cell level, infer the amino acid sequences encoded by the RNA, predict the protein structures of those amino acid sequences and simulate the protein-protein interactions within the target cell. The whole pipeline is still nascent and much work remains to be done, but the field of molecular biology is starting to look less like a collection of disparate, esoteric disciplines and more like the foundation of an entirely new era of medicine.
And there it is: your helicopter view of what’s hot in medical AI for the new year.
Which gives me just enough time to leave you with this parting thought:
Look to the future now, it’s only just begu-u-un…