21 – ROC Curve

Here’s our curve for Carcinoma. You find that in certain regimes, it’s 100 percent specific, 100 percent sensitive, and it’s only the little blue area on the top right that I are misclassifications. And there, we can now set thresholds to emphasize the detection of cancer or the reduce the cost of false diagnosis in the lab. Now, our very first doctor is this point over here, and it’s shocking to see how much fat is below or the blue curve. In fact for the same specitivity, he will miss three times as many cancers as our normal program is. That is quite significant, in my opinion. We then decided to test even more doctors and here’s a test set of 25 dermatologists all in the same data. And what you find is, there is a whole bunch of people that are willing to send people to the lab but still miss significant number of cancers where our evidence is effectively perfect. What’s also shocking about the result to me, personally, is the spread. There are some that are aggressively sending skin samples to the lab and have a good chance of finding your cancers, and others probably laugh by the endurances who don’t send few of the samples to the lab, and therefore, reduce diagnostic costs. Who knows? But you can see from this diagram, whether your personal skin cancer’s detected, depends a lot on which doctor you choose. And since this kind of data is never made public, we don’t even know which doctor are more specific, measure more sensitive to cancer. They appear the safer melanomas, the more important class, and yet the results are left with a conventional image. On the right side, results taken with a special instrument called a dormoscope, and you can see again the enormous spread in all these cases, the green dot is the mean of all the doctors. So, we went to all doctors and took the mean vote by the doctors. You would, in all cases, feel significantly worse than if a neural network looks at your skin.

%d 블로거가 이것을 좋아합니다: