A study on the selection criteria of medical images requiring human annotation based on uncertainty in semantic segmentation during active learning

Abstract

To build semantic segmentation models for interpretation of medical images, doctors need to make supervised images. Active learning is adopted to decrease annotation cost. However, little is known about the effective allocation of the supervised/un-supervised data during the first and the later training. We investigated the effect of the fluctuation of its allocation using 1463 Intravascular ultrasound images and discussed effective strategies of active learning. We first built three models which used 107, 359, 723 images respectively. Three models were built using 400, 700, 1000 supervised images and images sets predicted by the first models. The result suggests that the same accuracy is reached regardless of the number of the images at the beginning. We concluded that it would be efficient to start training even from a small number of supervised images and build following models using annotated images with higher uncertainty and predicted images.