Semi-supervised medical image segmentation of bladder tumors based on supervised branches and uncertainty estimation

Supervised learning has excellent segmentation performance in bladder tumor segmentation, but it relies on a large amount of labeled data. To learn the features of bladder tumors from limited labeled data and obtain accurate segmentation results, we propose a semi-supervised segmentation method for bladder tumors, UDS-MT. The method consists of a Mean Teacher network and a guided branch, which respectively undertake the tasks of segmentation prediction and supervision of prediction results. Mean teacher network uses the exponential moving average (EMA) mechanism to update the teacher network parameters to achieve fine-grained capture of the target shape; the guided branch uses uncertainty estimation to filter out pixel blocks with high confidence to obtain more reliable masks, and it suppresses overfitting on some certain extent. In addition, we propose a defend loss term that only calculates the loss for pixels with high prediction confidence of the model, thereby improving the reliability of the pseudo-label. After evaluation on a bladder tumor clinical medical image dataset, when the labeled data is limited to 15%, the Dice coefficient of the network segmentation target shape can reach up to 80.04%, which is at least 2.81% higher than other methods.