4 ms·

There was some interesting work recently published in nature on augmenting therapy selection: https://www.nature.com/articles/s41591-021-01359-w https://www.na

by new299 5y ago

There was some interesting work recently published in nature on augmenting therapy selection: https://www.nature.com/articles/s41591-021-01359-w https://www.nature.com/articles/s41591-021-01359-w “Overall, 89% of ML-generated RT plans were considered clinically acceptable and 72% were selected over human-generated RT plans in head-to-head comparisons.” This seems like it could be a way forward. Where AI is used to propose alternative and improve patient outcomes.

boleary-gl 5y ago
That's the way it is used today - for instance in mammography there is Computer aided detection (CAD): https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1665219/ https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1665219/ That's been in use for some time. But like many parts of radiology it really only can be a second look tool that as you mentioned proposes alternatives or suggests things. The false positive rate for CAD is substantially higher than for humans because of the human ability to see symmetry and patterns in very diverse tissue sets like one sees in screening mammography. And the nature of screening tests like mammography means that actually percentages like "89%" isn't really good enough. You have to be more specific and sensitive than that to have a successful program, and I'm not sure that ML will ever be able to get there...there's a lot of experience and human intuition involved at some point that would be hard to replicate...and I know that because people have been trying to do that for decades.
- ska 5y ago
  The comparisons are pretty tricky to do right, especially with systems that have been trained with the assumption that they are operating as "a second check". For what it's worth, that language was popularized by the first such system approved to market by FDA, in the mid-late 90s. It had, amongst other things, a NN stage. Even at that time, such systems were better than some radiologists at most tasks, and most radiologists at some tasks - but breadth was the problem, as was generalization over a lot of different set ups. I think this is more a data problem than an algorithmic one. With something as narrow as screening mammo CAD (very different than diagnostic), it's quite plausible that it could become a more effective "1st pass" tool than humans on average, but to get there would involve unprecedented data sharing and access (that 1st system was trained on a few thousand 2d sets, nothing like enough to capture sample variability)

by new299 5y ago

boleary-gl 5y ago

That's the way it is used today - for instance in mammography there is Computer aided detection (CAD): https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1665219/ https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1665219/ That's been in use for some time. But like many parts of radiology it really only can be a second look tool that as you mentioned proposes alternatives or suggests things. The false positive rate for CAD is substantially higher than for humans because of the human ability to see symmetry and patterns in very diverse tissue sets like one sees in screening mammography. And the nature of screening tests like mammography means that actually percentages like "89%" isn't really good enough. You have to be more specific and sensitive than that to have a successful program, and I'm not sure that ML will ever be able to get there...there's a lot of experience and human intuition involved at some point that would be hard to replicate...and I know that because people have been trying to do that for decades.

ska 5y ago
The comparisons are pretty tricky to do right, especially with systems that have been trained with the assumption that they are operating as "a second check". For what it's worth, that language was popularized by the first such system approved to market by FDA, in the mid-late 90s. It had, amongst other things, a NN stage. Even at that time, such systems were better than some radiologists at most tasks, and most radiologists at some tasks - but breadth was the problem, as was generalization over a lot of different set ups. I think this is more a data problem than an algorithmic one. With something as narrow as screening mammo CAD (very different than diagnostic), it's quite plausible that it could become a more effective "1st pass" tool than humans on average, but to get there would involve unprecedented data sharing and access (that 1st system was trained on a few thousand 2d sets, nothing like enough to capture sample variability)