Few-sample Multi-organ Abdominal Image Segmentation with Mean Teacher Model
Abstract
Medical segmentation is a significant task since it provides valuable information for diagnosis. In the recent years, convolutional neural networks have achieved great success in this field. However, the number of medical images is often limited which cannot support large networks to be trained. Thus, overfitting is much more common in medical images than other tasks. Also, to get annotated images is very costly and time-consuming. We stimulate an extreme condition where the number of the sample is so limited that we adopt a Mean Teacher Model to avoid overfitting. We build two models - student and teacher, with same structure and alternatively trained. We apply consistency loss to update the parameters of student and use Exponential Moving Average to compute parameters of teacher from student model. All code can be found in https://github.com/cpystan/MT-Model.