Long ago when I was a kid and my dad was aggravated, he’d sometimes tell us in jest, “You can be replaced by an adding machine!” What he was referring to was the fact that he had witnessed a dramatic transformation from his time in engineering school using slide rules to his, at the time, rare “brand-new” 1970’s Texas Instrument calculator. The same radical transformation is about to take over parts of medicine, but this time the disruption will be far greater because we have finally learned how to mimic the processes in the human brain. The first medical specialty that’s about to go the way of the dinosaur is diagnostic radiology. A new study out this week confirms what some already knew, that machine learning can read images far quicker and far cheaper with the same accuracy as a human radiologist.
All of the disruption we have seen in electronics and computer hardware and software has been based on humans programming systems to tell them what to do. This works well for things that can be easily fit into instruction sets, like math. It also works well for things like chess or searching massive databases. Where it breaks down is in real-world tasks. So while a computer that can search a billion records and find something in a second via a keyword seems smart, it fails miserably at something as simple as identifying a cat from a dog in a picture. Why? Even a shadow on the cat or dog will mess up any algorithmic approach you try to create for recognizing cats versus dogs.
Enter artificial intelligence—aka AI, neural networks, and machine learning. These are computer programs, and increasingly hardware to match, that work more like the human brain. They recognize patterns in large datasets and tie those patterns to some outcome. In other words, once you have programmed the initial program and how it’s set up, it “learns.” In the case of cats and dogs, you feed the machine tens of thousands of pictures of cats and dogs and tell it which is which. After that, it can recognize cats and dogs.
An interesting side note is that we have no idea in the traditional sense how this works. At a basic level, we understand that the system is set up like the brain with many connections that are either excited or inhibited, but we can’t deconstruct this in the same way a computer program filled with instructions can be torn apart, observed, and put back together. The patterns of the neural activations can be so complex that they defy human understanding. Also, increasingly, the patterns these systems find in large datasets also defy human understanding. For these tasks, the machines are capable of seeing patterns in the data that we humans are simply too dense to interpret.
The new study was performed using what’s called a deep neural network that was trained on more than fifty thousand hip radiographs with and without fractures. In the end, the system was able to detect hip fractures with human-level accuracy and false negative rates. However, it can read a thousand images a second for only the cost of a licensing fee, displacing hundreds of radiologists.
Should radiologists begin looking for a new job now? It will take some time for these systems to be complete enough to read all areas of the body and add in other modalities like MRI. It will also take time for insurers to begin reimbursing for these types of reads and for malpractice carriers to absorb what will happen when the machine misses a fracture. However, the handwriting is on the wall. These systems will likely end up with better accuracy, read much faster, and will cut out an entire layer of costs for insurers.
My guess is that we’ll see massive unemployment in diagnostic radiology by 2030. Hence, if I were a radiologist in my 50s, I wouldn’t be too worried. If I were a radiologist in my 30s, I’d be using this time to retrain into interventional radiology performing procedures. Finally, as a resident, I would not choose diagnostic radiology as a profession.
We have been using a neural network for some time to predict which patients will respond to a same-day stem cell procedure for knee arthritis based solely on some demographics and the chemical content of the knee synovial fluid. Right now our experimental model is 91% accurate. Hence, it’s better than physicians at picking out who will respond and who will fail. We will soon move this experiment to more Regenexx network sites outside of Colorado.
The upshot? AI is about to change society as a whole, and medicine won’t be left out. In fact, in many ways, it may provide the biggest societal “bang for the buck.” Fields in medicine that are diagnostic, like radiology, will be the first to become automated. In addition, we’ll also see other fields, like parts of internal medicine and cancer care, become automated, but these situations will be more likely doctors working alongside AI systems. So my dad with his fancy TI calculator in the ’70s was right: at some point, we will all be replaced by an adding machine!
About the Author
Christopher J. Centeno, M.D. is an international expert and specialist in regenerative medicine and the clinical use of mesenchymal stem cells in orthopedics. He is board certified in physical medicine as well as rehabilitation and in pain management through The American Board of Physical Medicine and Rehabilitation.…