R2Net: Recurrent Recalibration Network for Medical Image Segmentation

We analyze the prediction confidence heat map of current state-of-the-art models in medical image segmentation. We find that low confidence usually comes with a high probability of error. Based on this observation, we propose a novel Recurrent Recalibration Network (R2Net) for medical images segmentation. Specifically, the decoder will forward multiple times in a recurrent manner while the encoder forward only once to extract features during training time. Except for the first time, each decoder input features at time t will be recalibrated by the confidence heat maps which were output from the decoder at previous time t −1. As a result, R2Net can learn to focus on the regions which are error-prone to segment. The recurrent manner makes the feature reusable and saves the parameters. We validate the proposed models on two benchmark datasets: multiorgan segmentation and cardiac organ segmentation datasets. The experimental results show that our Recurrent Recalibration scheme consistently improves the segmentation performance of various models while preserving computational efficiency.