Imaging and Information in Radiology

Can informatics fill Radiologists shoes?

UK Shortage of Radiologists

Recently the Royal College of Radiologists expressed concerns over the shortage of radiologists in the UK;

I can't overestimate how worrying it is. "I do really feel the entire service will collapse if something isn't done about training more radiologists in the UK to fill all the vacant consultant posts, so we have to get imaging properly staffed - and right - and enough radiologists trained to make up this deficit.

The Promise of Technology

The promise of technology is that through clever processing (whether that be analytics, or machine learning) that information technology and systems can somehow take away some of the burden, and some of the mundane tasks, from expert jobs. This can help reduce the workload of the expert human in the chain of work.

Interestingly, I remember discussing this kind of analytical approach 20 years ago, before the shortage of radiologists became so acute. Back then my contention that signal processing could spot micro-calcifications better than a tired radiologist was rather dismissed - now these ideas are becoming more mainstream: but they are not yet wholly accepted, and nor are they yet 'the complete package. If technology is to fulfil its promise, there needs to be some serious commercial consideration for how this gets into mainstream use and right now I see more commercial barriers than technical ones.

Commercial Barriers to Analytics and AI in Radiology

The analytics and AI world for radiology is quite fragmented. Back in the old days, before we had the interoperability standards of DICOM for imaging, all imaging companies hid their image formats in a wrapper of intellectual property. Opening up this imaging world using DICOM meant that the market was free to pick and choose which platforms they viewed images on (though there are still some limitations here - see my discussion on cross-site reporting). Now Analytics packages and machine learning is bolted into the various imaging platforms (mainly the big PACS). And, indeed, why not? The imagnig companies have invested heavily in the research to produce the packages. The trouble is - no one provider has it all. The existing solutions are bitty and partial - and it is very difficult to compare what you are actually getting between vendors.

Two important things need to happen

  1. A set of regulated procedures to ensure the efficacy of the analytics and machine learning solutions on offer. Just as NICE performs this task for drugs in the UK, so we need a rigorous management to ensure the claims made by suppliers do indeed match the reality.
  2. An open interoperability platform to allow analytics and AI to be incorporated into wider use. The American College of Radiologists is indeed starting this off (just as they did with DICOM in the early days) but adoption by suppliers is slow. Purchasers need to recognise that suppliers will always tend to protected IP models (this is not a bad thing, just market forces). If users and healthcare providers want an open platform approach, they need to drive this agenda with their suppliers.

The reason this needs to happen is to open up the market. In the UK there is not a lot of money being invested in the world of radiology - in order to fund the research in the supplier companies, money needs to be paid, but individual use cases and applications are too small and insignificant to release the cash (benefits) from the purchasers. Opening up the market will allow freer interchange of services and make overall improvements of value.

How can technology really help?

Actually, I will extend this question to be, "how can technology help without undermining the expertise of the radiologist?" This is really important - despite the hype, we are a million miles away from a machine taking over the job of a radiologist. We need to prop up the world of radiology which is creaking under an increasing workload and not undermine it. One very important problem to recognises is a problem that is common to the computerisation of healthcare across the board: that a lot of input from users doesn't benefit 'the input users', but rather benefits other users downstream of them. For example, the relative merits of image annotation and mark-up compared to simply dictating it in a report is the subject of many a scholarly radiology article. What is the due consideration to be given to this subject? Are you handing information down the line, or a puzzle for people to put together from the bits of information you provide? Could a sufficiently trained application place the annotations and ROIs on an image based on the contents of a structured or natural language processed report?

Every way you turn in this subject you uncover complexity. The only way this kind of technology will be adopted will be if it reduces complexity and is thoroughly reliable. Easy to say. So what are my top tips for analytics and machine learning that will be able to have an effect on radiology?

  1. What's the difference between Computer Aided Diagnostics and Computer Assisted Decision Support? Not a lot in my view. CAD came to the fore (especially in Mammography) in the early 2000s. Peolle understood it had its proper place in a reporting or screening workflow. This is a sure fire growth area since this kind of application really supports a radiologist.
  2. Advanced 3D and automated segmentations - perhaps not necessarily for Radiologists' day-to-day work, but to a way to open out imaging for other clinical specialists in support of their work - particularly in orthopaedics.
  3. Functional imaging. Anywhere where a measure has to be made (like an ejection fraction) - judgement of where to place ROIs can be more consistently handled by machines than by disparate people (or even the same person, comparing their performance over a session from 'new and fresh' at the start of a session and fatigued by the end of it).
  4. Workflow improvement tools - measuring efficiencies and feeding them back into continuous business improvement.