AI does not increase productivity in radiology - yet
It is a well-known paradox that technology adoption, not only related to AI, is leading to a J-curve shaped change in productivity. Easy example: Before the introduction of typewriters, reading written text was painful. It was hard to read other people's handwriting and after some time, the ink bleached and made texts unreadable. Then, the typewriter was invented. For us in the 21th century, it is an ancient machine but back in the 19th century people had never used such a device before and didn't know how to type on a keyboard efficiently using 10 fingers. As a consequence, writing text with a typewriter was less productive than handwriting initially, i.e. the new technology led to a decline in productivity. However, by the time, people learned to type efficiently, and the advantages of typewritten text outweighed the pain of learning to type with 10 fingers. Net productivity increased a while after introduction of the new technology.
I think that the very same logic can be applied to AI in medicine. The J-curve can be partitioned in three phases:
Phase 1 - Status quo 📠
Baseline, i.e. before AI-adoption. Most physicians write medical reports as prose, lab results and measurements from separate data silos are manually combined in written reports. This is the status quo as of 2021 for most doctors and a non-negligible minority is still using fax machines...🙄
Phase 2 - Transformation 🚧
The party begins. The first certified machine learning-based products that are supposed to support physicians have entered the market. Initially though, a doctor's life does not seem to be easier but even more complicated: It is not enough to write a report and communicate one's findings with patient and referring physician. For an early-adopter radiologist, there are not only dozens of images but also PDF-reports - automatically generated by AI - in the PACS system. These reports need to be interpreted in the clinical context, which is additional work. Ah, and what about quality control for the input data to the AI? Is that particular case today even within the specification of the product? Has our IT-department already set up the GPU-workstation for AI-tool X by vendor Y? The list of new questions that have not been there before the advent of the new technology goes on and on....
But things are about to change...
Phase 3 - Increased productivity 🚀
If you looked closely at my illustration above, you realized that phase 3 begins when the productivity is at break even as compared to before introduction of the new technology. In other terms: At break even, we are as productive as at the end of phase 1, but the underlying tools and workflows are based on more modern technologies. This is only the beginning, though. As products mature and we are getting more used to them, we can increasingly benefit from the underlying technology. What does this mean specifically? Let me break it down into three aspects:
- Technical infrastructure for AI: As of today, it is quite common to set up dedicated hardware for every AI product that a clinic wants to use. A better solution are abstraction layers, i.e. "app stores" that provide access to a whole collection of AI-tools using a single coherent interface. This will drastically reduce upfront and maintenance cost and simplify billing for various AI-tools. With DICOMweb, FHIR and cloud computing, all technology is already available. Apple has proven that the App Store business model is highly successful. So, I think it is just a matter of time until that part of the problem is solved.
- Workflow-efficiency: In order to allow a very simple and efficient use, AI tools must be tightly integrated into 1) image viewing and 2) report writing workflow. Many AI-tools will be used for measuring, which is good because this may reduce intra and inter-rater variability. However, as more measurements are used in the report, the report structure will be even more important for an efficient AI integration. Therefore, I expect that DICOM-viewer and structured reporting software will merge into one single coherent application. This will also allow a more interactive integration, e.g. interactive AI-based segmentation for calculating a volume and directly pasting that volume in the structured report. This will not only speed up reporting but also improve reporting reproducibility. Finally, as most AI software will run in the cloud, modern web-based interfaces like DICOMweb and FHIR will finally become mainstream, which paves the way for user-friendly responsive web-viewers.
- Reproducibility: After spending the last five years of my life in radiology AI/computer vision research, I am pretty sure that this will be a hard nut to crack. Sure, there are many papers that solve a very narrow and highly specific question with high accuracy. However, delivering sufficient accuracy across scanner vendors, field strengths, reconstruction algorithms, and real-world patient populations is orders of magnitude harder than writing a paper. I hypothesize that we don't even know very well how big of a problem this is, because certified AI tools are still having very small market shares. Most tools have not been evaluated prospectively in multi-center studies as recently reported in Nature Medicine. Since the protocols are highly standardized, neuroimaging applications are probably best suited for 1) quantifying reproducibility problems and 2) developing methods to improve reproducibility.
Where are we today?
In short, we are at the beginning of phase 2.
Early-adopters like my colleague Christoph Agten can already use the first certified tools in clinical routine. But there are huge challenges ahead and solving these challenges offers massive opportunities for startups who make AI easier to use for radiologists. This trend is even showing in the market already: Three years ago, AI startups were all about diagnosing pathologies. Today, companies focusing on workflows are gaining traction: Viz.ai, who are already focusing on improved workflows for stroke care have recently raised 71 million USD to expand their business to other verticals. deepc who aim to develop efficient infrastructure ("operating system") for radiology AI raised a mid seven-figure sum.
To conclude, it is still very early in the game and I expect early adopter radiologists to suffer quite a bit in the next 24 months. But there's hope: Many smart people are already working on making AI workflows smoother for clinicians and improving productivity.