Hacker News Reader

Comment by roentgen | Hacker News Reader

I'm not sure if you're being serious or not, but I'll answer.

I happen to think real AI will not happen in my lifetime, if ever. This is not an uninformed opinion, I have MS in Neuroscience, and I'm a programmer. I'm sure others reading HN will disagree.

It will take nothing less than a full artificial intelligence to do what I do.

If you want to read more about early attempts, there is something called "Computer Aided Detection" that is used when reading mammograms. It is awful, and I'll let you in on a secret. Most places that CAD their mammograms do it so they can charge more, not because it helps.

View on HN

You mentioned in an earlier comment about a CT scan being 6k+ images. I'm assuming these are cross-sectional slices or something similar?

I know a guy who's working for one of the big CT manufacturers (Toshiba, I think) working on 3D image-registration and partial analysis. In essence, lining up the features from different slides and sewing them into a 3D model.

One of the big features being looked at was automatic labelling of certain gross anatomical features - I seem to remember something about cardiac veins needing to be marked as part of a scan report? By mostly automating this, it saves time for the user so they can just check they're correct, maybe move them if not, rather than having to do all the work themselves.

I can see this sort of work being much more useful, and happening in a much shorter term than any sort of AI "You have cancer" system. I wonder how you feel about that sort of thing?

I use this software when we perform a CT scan looking at the coronary arteries. Under perfect conditions it can correctly idenitfy the coronary arteries, and subtract away the rib cage, heart, and lungs.

It sounds like you already understand what it does. It helps me read a study quicker by automatically processing the data, a step I used to do by hand. Many times the processing fails due to an artifact while scanning ( patient moves, ectopic heart beat, poor contrast injection timing, variation in anatomy), and I need to process it manually.

When it works well it is very helpful, but to be clear, it does not interpret the studies. I think many people replying in this thread don't understand the enormous complexity to accurate radiology reporting.

Reply on news.ycombinator.com