Semi-Supervised Deep Learning using Consistency-Based Methods for Segmentation of Medical Images

Sammanfattning: In radiation therapy, a form of cancer treatment, accurately locating the anatomical structures is required in order to limit the impact on healthy cells. The automatic task of delineating these structures and organs is called segmentation, where each pixel in an image is classified and assigned a label. Recently, deep neural networks have proven to be efficient at automatic medical segmentation. However, deep learning requires large amounts of training data. This is a restricting feature, especially in the medical field due to factors such as patient confidentiality. Nonetheless, the main challenge is not the image data itself but the lack of high-quality annotations. It is thus interesting to investigate methods for semi-supervised learning, where only a subset of the images re- quires annotations. This raises the question if these methods can be acceptable for organ segmentation, and if they will result in an increased performance in comparison to supervised models. A category of semi-supervised methods applies the strategy of encouraging consistency between predictions. Consistency Training and Mean Teacher are two methods in which the network weights are updated in order to minimize the impact of input perturbations such as data augmentations. In addition, the Mean Teacher method trains two models, a Teacher and a Student. The Teacher is updated as an average of consecutive Student models, using Temporal Ensembling. To resolve the question whether semi-supervised learning could be beneficial, the two mentioned techniques are investigated. They are used in training deep neural networks with an U-net architecture to segment the bladder and anorectum in 3D CT images. The results showed signs of promise for Consistency Training and Mean Teacher, with nearly all model configurations having improved segmentations. Results also showed that the methods caused a reduction in performance variance, primarily by limiting poor delineations. With these results in hand, the use of semi-supervised learning should definitely be considered. However, since the segmentation improvement was not repeated in all experiment configurations, more research needs to be done.