Abstract
One way of collecting huge ultrasound dataset is by taking ultrasound videos. But one drawback of using video is that many frames look alike and without proper selection, feeding the visually similar images as input for the training may lead to overfit to those ultrasound images. And to avoid such an effect, we need to define the similarity metric between ultrasound images and analyze the relationship between similarity metric and overfitting effect. In this research project, we try to suggest a number of selecting algorithms and assess the affect of the algorithms on the training result.