I made the mistake of procrastinating on something more meaningful by reading a variety of random commenters on issues related to radiology. One type of flawed thinking stuck out: the all-or-nothing fallacy.
For example, as it pertains to artificial intelligence, the argument often goes, “AI will never replace a human in doing what I can do, and therefore I can ignore it.” Or, “I put a radiology screenshot into a general-purpose LLM and it was wrong” or “our current commercially available pixel-based AI is wrong a lot, ” therefore, “I can ignore the entire industry indefinitely based on the current state of commercially available products.”
Leave aside the potentially short-sighted disregard for this growing sector because of its obvious and glaring current shortcomings. Even the current state of the art can have an impact without actually replacing a human being in a high-level, high-training, high-stakes cognitive task.
For instance, let’s say the current radiologist market is short a few thousand radiologists–roughly 10% of the workforce. Basic math says we could:
- Hire 10% more human beings to fill the gap (difficult in the short term)
- Reduce the overall workload by 10% (highly unlikely)
- Increase efficiency by 10%
The reality is, it doesn’t take that much magic to make radiologists 10–20% more efficient, even with just streamlining non-interpretive, non-pixel-based tasks. If only enterprise software just sucked less…
We don’t need to reach the point of pre-dictated draft reports for that to happen. There’s plenty of low-hanging fruit. Rapid efficiency gains can come from relatively small improvements, such as:
- Better dictation and information transfer. When dictation software is able to transcribe your verbal shorthand easily (like a good resident), radiology is a whole different world.
- Real summaries of patient histories.
- Automated contrast dose reporting in reports.
- Summaries of prior reports and follow-up issues (e.g., “no change” reports where previous findings are reframed in the customizable style and depth).
- Automated transfer of measurements from PACS into reports with series/image numbers.
- Automated pre-filling of certain report styles (e.g., ultrasound or DEXA) based on OCR of handwritten or otherwise untransferable PDFs scanned into DICOM.
These tasks, as currently performed by expensive radiologists, do not require high-level training but instead demand tedious effort. Addressing them would reduce inefficiency and alleviate a substantial contribution to the tedium and frustration of the job.
Anyone who thinks these growing capabilities–while not all here yet, nor evenly distributed as they arrive–can’t in aggregate have an impact on the job market is mistaken. And if AI doesn’t develop fast enough to prevent the continued expansion of imaging responsibilities to non-physician providers, the radiology job market will be forced to contend with a combination of both factors, potentially leading to even more drastic consequences.
When you extrapolate a line or curve based on just two data points, you have no real idea where you started, where you’re headed, or where you’re going to end up. Just because you can draw a slope doesn’t make the line of best fit meaningfully reflect reality or extrapolate to a correct conclusion.
Don’t fall prey to simple black and white thinking.