[bugsbunnyak]

There's a small army of PhDs doing this right now, NIH-funded and in commercial research shops.

A radiologist still must look at every single slice for liability reasons for the foreseeable future. That's the real time cost.

I work for a neurosurgeon in image-guided surgical planning research. One of the challenges is segmentation (labeling) of target areas to use in navigation. (radiologists generally don't do this, for various reasons). I've used some of the best commercial software, and seen some of the top research algorithms. With these, for the `easiest` tumors, we still have to semi-manually choose the region on every 3rd or 4th slice. The best algorithms will interpolate the other slices based on essentially fitting along a levelset. For a typical tumor, it can take 20-40 minutes to do this task - using the best available software!

This is `not` radiology, it's image labeling. It's orders of magnitude simpler than radiology.

There are some promising techniques to, for example, automate detection of changes in volume of some radiographically questionable area (after manual labeling for the first scan). At best, this will add information with no extra time cost.