An active learning model based on image similarity for skin lesion segmentation

Medical image segmentation requires experienced doctors to provide accurate segmentation results as ground truth, since it involves plentiful medical knowledge. The deep learning-based model requires a large number of medical images and ground truth, during the training stage, which brings a heavy workload to the doctor. Therefore, we design an active learning based on image similarity framework for selection strategy to select more representative images for doctors to obtain the ground truth, which saves their time in segmenting similar medical images. Eventually, these representative images will serve as training sets to train amore accurate segmentation model. In this paper, we propose the active learning-based segmentation (ALS) model for medical images, in which we use the mean structure similarity index measure to measure how similar the two images are. Then we select some highly similar images, since medical image segmentation is different from semantic segmentation, there is no interference from similar targets, so we do not need to cull interfering images when segmenting medical images. Finally, we reduce the amount of data to 75%, 50%, 25%, and 12.5% to validate our ALS model. Even a small amount of image through ALS model selection as a training set, and the final segmentation model of training also has a good performance. The active learning model can reduce the cost of labeling and obtain accurate segmentation results through experiments, and even better than the segmentation results with a large number of highly similar data.