[yoden]

The problem with your thesis is that many radiologists are paid per read. They're essentially contractors. These independent radiologists would gladly pay 100k/year for an AI that could do their jobs. Even if the tool didn't do diagnosis, just made them 50% faster, they would gladly pay a huge amount of money.

Contrast this with the field of Radiation Oncology. There are already AI auto-contouring solutions that you can buy today. It's a better fit given the limitations of AI:

1) Contouring can take hours per patient, so more value is gained by automating it

2) Normal structures are, well, normal; that is that they are not the diseased part of the patient. If the AI can't quite figure out where the kidney is and fills it out with an average looking kidney, hey that's probably an alright guess!

3) The AI doesn't have to be perfect; dosimetrists and docs can quickly check and update the result if the algorithm fails. Everyone is already familiar with this workflow due to the atlas segmentation algorithms that were state-of-the-art before.

4) Similarly, the output is easy to understand. If the AI did something crazy because of a failure to generalize, it will be obvious. Not so for a diagnostic AI.

5) Clinicians have the final say on which contours are used for treatment. The liability for contour assistance is lower for the software vendor.