An Active Learning with Two-step Query for Medical Image Segmentation

Active learning (AL) is a special machine learning method, in which a learning algorithm is designed to query the user to obtain the desired label interactively. It has been widely used in the area of image classification and not effectively introduced to the field of segmentation so far. However, the cost of segmentation’s annotation is dramatically greater than the classification. It’s more useful to apply active learning in the segmentation task. Thus we bring forward a novel two-step query method in this paper. First, we design a fusion of multiple estimate models to calculate sample complexity as the middle variable. Second, we define a function of sample complexity and potential value, which can also be interpreted as activation and help the interactive query collect more ponderable samples. Through the proposed two-step query method, we progressively optimize the model’s capacity for sample selection and object segmentation. Experimental results of the bladder segmentation task show that our method outperforms random sampling about 12% relatively using the same number of data.