I need some help regarding the selection of the sample size. I am working on medical image segmentation and I have to measure the accuracy of segmentation using the Jaccard index. The issue I am facing is that my data (images) collection is from different sources and for all of them, the ground truth is not available. I have five classes in total but I only have ground truth for three class. Even for these three classes, I have ground truth only for some of the images. What I found from different articles is that I have to select a sample size and on the basis of the sample size I can generalise for the whole dataset. For sample size selection, I need error rate and Confidence interval. Now my questions are: Q1) The standard confidence interval is usually 95 %, Is there any by the default value for error rate if don't have any idea what our error should be? Q2) Can I assume that if my segmentation technique is performing well on three classes, It will also work on other two classes for which I don't have ground truth? Based on statistics, what assumption can I make for full dataset based on my problem? Any guidance will be much appreciated.
Thanks in advance