12 – Validating The Training

To validate our results, we constructed an independent test set, and we’ve worked really hard to clean up the data. The data was, in many cases, had duplicates, we didn’t want just split trading and test and there’s a chance that the same image occurs in training and testing. It had some data markers, some dermatologist had put markers on specific classes. Like they picked yellow markers for size determination. And of course, we didn’t want to take an interval to find yellow markers, we wanted to find cancers so a way to remove those images and clean them up. But after the clean up, we were able to take a subset of the images to be testing set, and as you always do to validate your results. And interestingly enough, the very first time we did a three way classification between melanoma’s, carcinomas, and P nine we, got a 72 percent accuracy in finding cancers. And, we felt that was significant. So we decided to go to the School of Medicine and open the door of the real dermatologist. And the person whose name I named here, was willing enough to look at the same images as we used for testing and was 65 percent correct. A whooping 6.4 percent less than us. That is significant in cancer detection. So we knocked on the door of a second dermatologist and asked the same question. He was willing enough to do it, and we got to 66 percent correct classification rate. So we had all of a sudden, in knowing that rock that, if looking at images seemed more accurate than two board certified Stanford level of interest in finding skin cancer, that got our attention. So now we decided to do a real experiment.

%d 블로거가 이것을 좋아합니다: