Multi-organ Segmentation Algorithm Based On Partially Labeled Data

Multiple organ segmentation has been widely used.But annotating multiple organs in medical images is both costly and time-consuming? therefore,existing multi-organ datasets with labels are often low in sample size and mostly partially labeled,that is,a dataset has a few organs labeled but not all organs.While the current organ image segmentation method mainly relies on the distribution of existing data.The larger the amount of data is,the more generalized the data distribution will be.Therefore,an algorithm that can use multiple data sets from different sources to jointly train segmentation is needed.The research on the multi-organ segmentation algorithm of partial label data is carried out.The existing multi-organ segmentation algorithms is investigated,including a variety of technical route methods based on multi-atlas,based on prior knowledge,and based on deep learning.The research of semi-supervised learning algorithms in various scenarios such as classification,recognition and segmentation is investigated.The development status of deep learning in medical image segmentation is investigated,including commonly used deep learning models,deep learning modules,and data processing methods.How to learn a single multi-organ segmentation network from a union of such datasets is investigated.To this end,two types of novel loss function is proposed,particularly designed for this scenario:(i)marginal loss and(ii)exclusion loss.Because the background label for a partially labeled image is,in fact,a ‘merged’ label of all unlabelled organs and ‘true’ background(in the sense of full labels),the probability of this ‘merged’ background label is a marginal probability,summing the relevant probabilities before merging.This marginal probability can be plugged into any existing loss function(such as cross entropy loss,Dice loss,etc.)to form a marginal loss.Leveraging the fact that the organs are non-overlapping,the exclusion loss is proposed to gauge the dissimilarity between labeled organs and the estimated segmentation of unlabelled organs.Experiments on a union of five benchmark datasets in multi-organ segmentation of liver,spleen,left and right kidneys,and pancreas demonstrate that using our newly proposed loss functions brings a conspicuous performance improvement for state-of-the-art methods without introducing any extra computation.Besides,we also proposes a structural cross-sharing convolutional layer that can learn adaptive shared features from multiple data sets with different standard from a structural point of view.Using the image features learned by the respective branch network of each task,the self-convolution between features is realized through cross-sharing convolution.Adapt to sharing,and finally achieve the goal of parallel learning with multiple data sets.In the liver and spleen dual-target organ segmentation experiments,it is shown that sharing features can further improve the segmentation effect.Through the two methods proposed from the perspective of loss function and network structure,using multiple partially labeled organ segmentation data sets to train the multi-organ segmentation network effectively improves the performance of the network.In addition,better results have been obtained in comparison with other methods that also use multiple partial annotation data at this stage.