[paper reading] multi task attention based semi supervised learning for medical image segmentation

[ Address of thesis ] [ Code ] [MICCAI 19]

Abstract

We propose a new semi supervised image segmentation method , It optimizes both supervised segmentation and unsupervised reconstruction objectives . Rebuilding goals uses an attention mechanism , Separate the reconstruction of image regions corresponding to different categories . The proposed method is evaluated in two applications ： Segmentation of brain tumors and white matter hyperdense areas . Our method trains on unlabeled and a few labeled images , Its performance is better than that of supervised training with the same number of images CNN And pre trained on unlabeled data CNN. In ablation experiments , We observed that , The proposed attention mechanism greatly improves the segmentation performance . We explored two multi task training strategies ： Joint training and Alternate Training . Alternate training requires fewer super parameters , And achieved better results than joint training 、 More stable performance . Last , We analyzed the characteristics learned through different methods , The discovery of attention mechanism helps to learn more discriminative features in the deeper layer of encoder .

Method

The idea of semi supervision in this paper mainly includes two aspects . The first is Multi-Task, That is to build an additional decoder $D_R$ To perform the task of unsupervised reconstruction , Is equivalent to Encoder-Decoder As a self supervised VAE; There are two ways of thinking about this unsupervised reconstruction , One is to directly reconstruct the original image itself , The other is to reconstruct the foreground and background of the original image respectively , This paper finds that the latter is better .

And as for this attention, It refers to the rough result of prediction (soft segmentation) It is not directly used as a pseudo tag , Instead, it multiplies the input image by a pixel ; In this case , The coarse foreground and background of the image can be extracted . This rough foreground and rough background need to be consistent with $D_R$ The prospect of reconstruction is consistent with the background , So as to achieve consistency constraints for semi supervised learning .