In an article titled “The Robot as Radiologist,” Dr. Douglas Green from Univ. of Washington acknowledges the rapid advances in computational image recognition and advent of IBM’s Watson. He concludes the commentary by taking solace in the fact that, at least for the time being, artificial intelligence is complementary rather than substitutional to human radiologists. I wholeheartedly agree. However, Harvard Business Schools gurus do not.
The Complexity of Machine Learning
The problem of “can robots replace humans” is very much like the “P ?= NP” problem in mathematics. Tasks that can be remarkably simple for a human to perform actually have no easy solution computationally: Facial recognition and speech are the prime examples. Extracting features from radiology images falls under the same category. In fact, radiologists use the phrase “Aunt Minnie” to describe diagnoses that can be instantly identified like one might recognize the face of one’s relative in a crowd.
The field of machine learning tries to tackle these hard problems by simplifying the task into patterns and by approximating a solution. Although the actual solution may be NP-hard, both facial and speech are recognition can now be solved using polynomial-time heuristics as long as you are willing to tolerate some error – i.e. a good answer most of the time rather than the best answer all the time.
Radiology interpretation could go the same way. Radiology informatics is laden with approaches using segmentation, pattern recognition, and Markov chains. The disagreement lies in (1) whether machines are capable of performing these tasks, and (2) whether the value-add of a radiologist ends at interpretation.
Machine Intelligence in Radiology
A Harvard Business School blog post last December discussed what may happen after robots take over human jobs. The logic goes like this:
- The “smartest” robots are already capable of performing at a level of IQ=115.
- Technology advances by Moore’s Law
- Whenever new technology (i.e. robots) gets installed, workers get laid off.
Although falling short of outlining the steps of how Skynet might dominates humanity, Davidow and Malone did cite radiology as an example:
An emerging field in radiology is computer-aided diagnosis (CADx). And a recent study published by the Royal Society showed that computers performed more consistently in identifying radiolucency (the appearance of dark images) than radiologists almost by a factor of ten.
Will Robots Replace Radiologists? (No)
There should be no doubt that algorithms will achieve some level of success. The underlying problem doesn’t lie in the robotics of it all so much as replacement and decoupling in the value chain.
Radiologists are not alone in the value chain of radiology. Christensen’s classic disruption theory posits that technological advances tend to decouple tasks ordinarily performed by highly trained professionals. The theory would suggest that as tasks of imaging processing, analysis, and interpretation become standardized, those same tasks can then be completed to the same level of quality either by outsourcing or by less-trained professionals.
The same reason why standardization yields more predictable outcomes is the same reason why standardization tends to be a precursor to commoditization. And it is the reason why modern radiology’s seemingly paradoxical approach of simultaneously pursuing standardized protocol and reporting (every exam should look the same!) and personalizing imaging to the unique individual patient (every exam is different!) makes sense.
Every protocol in radiology should have an option for real-time modification based on the individual needs. Every “standardized” diagnostic report should leave room for a radiologist’s free text.
Davidow and Malone might consider these added flexibility a way that radiologists “up the ante” by taking ownership of tasks of increasing complexity that cannot be fully emulated by an algorithm.