Deep Active Learning for Cardiac Image Segmentation

In recent years, the incidence rate of cardiovascular diseases have been increasing. Cardiac cine magnetic resonance imaging (MRI) is an important method to detect cardiovascular diseases. In the diagnosis of cardiovascular diseases, semantic segmentation of left ventricular cavity, left ventricular myocardium and right ventricular cavity of Cardiac MRI data is a very important step. Now many researchers have proposed different heart segmentation methods. However, these methods need a large number of labeled data sets, and the labeling of these data sets is undoubtedly time-consuming and laborious. This paper presents a deep active learning method based on entropy. In each step of active learning, a batch of unlabeled samples with the largest entropy are selected by using a deep supervision network and handed over to human experts for annotation. The model is trained iteratively until it reaches the desired performance. The results of the experiment show that the active learning method we proposed is obviously better than the random sampling method, and only a small amount of labeled data is needed to achieve the segmentation results achieved by training the model with all data sets.

Automated summary

The incidence rate of cardiovascular diseases has been increasing, with cardiac MRI being an important detection method. Existing heart segmentation methods require a large number of labeled datasets, which is time-consuming and laborious. This paper proposes a deep active learning method based on entropy, which is shown to outperform random sampling and only requires a small amount of labeled data.