Michael Lewis warned that baseball scouts were heading for extinction. 15 years later, MLB teams had hired 32% more of them.

Geoffrey Hinton said we should stop training radiologists.

The Mayo Clinic now has 55% more of them.

Those touting the technology are quick to make predictions about people. But the technology doesn’t decide what happens. How people use it does. To predict that, you need to understand two things.

First, the job. The inputs and outputs, the edge cases, the workflows, the incentives.

Second, and often overlooked, the cognitive predispositions of the people doing it.

When I was building AI for radiologists, the industry was focused on one kind of error: missing something on the image. But radiologists make at least a dozen others. They anchor on the clinical history. They stop looking after the first finding. Put an AI flag on the screen and they may trust it too much, or after enough false alarms, not enough.

Baseball learned the same lesson. Teams had more data than ever, but a hitter can only hold a thought or two in the batter’s box. The edge went to the coaches who knew which one to give him.

The technology didn’t replace those people. It ran into how they think.

Now predictions abound about the rest of us. AI will replace knowledge workers, we’re told, starting with entry-level jobs.

I don’t think the Big AI labs understand the job.

Their predictions are making the same mistake: they’re looking at what the technology can do, not at how people will work with it.

Midjourney just entered medical imaging.

The medical community is focused on clinical gaps.

The tech community is focused on scale and power.

Both may be missing the point.

The goal of Midjourney’s technology isn’t to augment current modalities or replace them.

That’s because current modalities are engineered and optimized to generate images for human eyes. Before the images are rendered on screen, several compromises have been made so that what appears on screen maps to anatomy a doctor recognizes.

Current imaging modalities trade signal for things like contrast, geometric fidelity, artifacts, image uniformity, etc. An image is, after all, subjective and qualitative, even if we hope that the interpretation it enables is not.

But the real potential of Midjourney’s approach is that it could acquire data optimized for machine analysis. We have to remember the deliverable in medical imaging is not images. It’s a report outlining medical findings that guide patient care.

Blood tests are optimized to extract a signal, reported as a number (cholesterol level, a troponin reading) and the doctor acts on that number.

The right question isn’t whether Midjourney’s images are as good as an MRI or CT scan. A machine will not need to recognize a liver to detect that something is wrong with it. Once we accept that premise, all of medical imaging can be re-imagined.

But it will still have to be validated, one marker at a time.

I track what AI does to humans in their workflows. Follow if that angle is missing from your LinkedIn feed.

For experts, AI has potential, but will remain useless until it is recognizes the cognitive factors at play.

Radiology makes a great case study. The initial promise that radiologists would be replaced was naive to the fact that radiologists do more than read (interpret) images and report findings. Yet, handing them algorithms was never going to be enough.

Under-reading, what AI startups promised to solve, is only one type of error. Here are the other types:

• Finding is seen but dismissed (complacency).
• Finding is seen but attributed incorrectly.
• Finding is seen but not recognized due to a gap in knowledge.
• Finding is seen but miscommunicated.
• Finding is missed due to image acquisition error or quality.
• Finding is seen but clinical history is missing or inaccurate.
• Finding is missed because the interpretation stopped prematurely after another finding was seen.
• Finding is seen but there is an over-reliance on a previous report.
• Finding is missed to to inattentional blindness.
• Finding is missed because of bias of how the patient is presenting.
• Finding is missed because of the anchoring bias of earlier information.

Everyone of these error types will interact with a radiologist using AI in new and different ways. For example, AI results can be the source of anchoring bias or a new source of premature closure.

Furthermore, AI will create new types of errors:

• Automation bias (Over-confidently trusting AI results)
• Attentional tunnelling (Focusing on AI results at the expense of other findings).
• Algorithm Aversion (When a radiologist broadly dismisses AI due to past mistakes).
• Incidentalomas (incidental by inconsequential findings).
• Alarm fatigue (false positives reducing the trustworthiness of all positive AI findings).
• Context bias (the radiologist and AI may be operating with contextually different information).

Engineers, developers, lawyers, truck drivers, financial traders, recruiters, insurance brokers, journalists, adjusters, teachers inspectors, etc. all have such lists of their own.

Every field is full of AI blind spots. Do you know which ones you’re solving for?

After its AlexNet moment, AI was quickly aimed at medical images in radiology, in the form of neural networks. It made sense to everyone outside the field and produced a few famous claims that radiologists would soon become irrelevant. But, every startup that sought to replace radiologists has disappeared or pivoted. It’s been, speaking broadly, a bumpy road. Medicine remains a human endeavour.

The technical aspects of image acquisition are fascinating. It is the product of decades of research and development of very bright engineers, physicists, developers, material scientists, etc. These decades of effort have optimised image quality for interpretation by a medically-trained human. This point is worth repeating. Years of effort have optimised image quality in a largely qualitative way for various clinical scenarios. These improvements have been aimed to improve the suitability of the images for interpretation by a human - often on the balance of compromises such as radiation or scan time reductions.

Are the most useful images for a human also the most useful images for a classification or other inference algorithm? That seems unlikely. Then, when will we optimise the image acquisition, perhaps the engineering of the scanners themselves, to the benefit of our algorithms instead of humans?

It appears steps are already being taken in this direction.

The current software delivery model in healthcare … simply will not serve current and future advances in medical image analysis.

Many say this is about to change. Digital imaging data is a veritable treasure trove for the quickly developing field of computer vision, led by recent advances in machine learning. It is a data-rich opportunity to develop and apply new algorithms with a compelling value proposition. Therein lies a true opportunity to increase accuracy, save lives, time, and of course, money. While some would suggest that there is an aggressive charge to replace radiologists, I would agree with those who suggest the pace will be more timid and with the result being a collaborative effort between man and machine. But perhaps not for the same reasons. There is a strong focus on the performance, accuracy and clinical relevance, as there should be. But my concerns are with respect to the current software delivery model in healthcare: it simply will not serve current and future advances in medical image analysis.

Fields which have been revolutionized by software have done so by enabling a tremendous pace of change. While the rate of innovation in medical image analysis has accelerated due to advances and access to software, the current paths to market for new software technologies limit adoption. Existing vendors with comparatively deep market reach have slow development cycles that match the long procurement cycles, while simultaneously locking technologies within proprietary frameworks.

Meanwhile, independent startups are able to focus on their technology and move much more quickly. However, the resulting implementation usually takes the form of a standalone application or web portal. These solutions are not typically integrated into current workflows of time-pressed clinicians and their installation may be met with resistance by those responsible for the IT infrastructure. Therefore, spreading the solution to many hospitals is difficult and clinical adoption is once again slowed.

These medical image analysis technologies, often simply referred to as algorithms, are not very broad in nature, but typically developed for very specific use cases. There are already dozens of applications in the works and we may see hundreds become available in the coming years as competing algorithms vie for the same clinical applications. Within the current framework, the majority would end up housed in silos, like much of medical data. This would restrict clinical flexibility and institutional mobility. And what about transparency? Will the field be able to compare proprietary algorithms to each other, including revised versions or those which continuously update themselves?

…the anticipated marriage of AI and radiology is unprecedented in medicine.

Currently, a common method to gain clinical adoption is for new entrants to partner with incumbents. While an improvement, given the above, we can see this as a sub-optimal solution. We have to remember that machine learning and computer vision all originated well outside the field of medicine. While applications from computer science (and many other fields!) have been improving healthcare for decades, the anticipated marriage of AI and radiology is unprecedented in medicine. Even large tech companies that had to date stayed at the periphery of healthcare are now joining the charge. At the other end of the spectrum, university researchers without access to direct clinical collaborators are still able to push the field forward with their contributions. Large open data sets will further this trend by removing one of the major barriers to the development of algorithms.

Based on all these reasons, we see that the current innovation-to-adoption framework for software in healthcare will not serve the influx of AI-enabled technologies. Between the proprietary frameworks of the incumbents and the immense hurdles faced by new entrants, we need to think about how we can remove these barriers to progress.

The introduction of AI-enabled technologies in radiology will result in new challenges and opportunities in the interpretation of medical images. It may appear as a primarily logistical challenge at first glance. But, once we delve into some specific aspects, we reveal implications which suggest that an entirely new delivery model must be conceptualized and built. And this can only be achieved with the participation of all stakeholders and the benefit of our patient population in mind.