Query-by-committee Self-paced Learning With Diversity Based Medical Image Segmentation

Recently,deep learning,especially for Convolutional Neural Network(CNN)has been widely use in nature image and medical image segmentation tasks.CNN can automatically extract features from images,and the automatic extracted features can be applied to corresponding tasks.Training a CNN based segmentation model often requires a large amount of annotated data.However,the number of pixel-wise annotated medical images are extremely small,which limits the application of CNN in medical image segmentation tasks.To fully use the advantage of the CNN in the application of image segmentation tasks,we proposed a self-paced learning with diversity(SPLD)framework by using Query-byCommittee(QBC)method.The proposed algorithm can select training samples from easy to difficult for model training so that to boost the performance of medical image segmentation models.The proposed SPLD algorithm imitates the learning process of a student obtains the knowledge,i.e.learning from easy to hard which guaranteeing the diversity of the knowledge and fully utilizing the limited annotated data,to improve the accuracy of the algorithm.The main work of this paper includes the following aspects:Firstly,we proposed the QBC based SPLD learning framework.The proposed framework employs the QBC algorithm to implement the data selection during the model training,and feeds the data from easy to hard sequentially to the model to improve the model performance.Meanwhile,the framework uses affine propagation to cluster the training data without supervision to guarantee the data diversity during the easy to hard data selection process.The proposed method implemented the diversity learning and prevented the model from optimizing to local minima.Sceondly,we prove the effectiveness of the proposed approach quantitatively and theoretically.For the quantitative analysis,by comparing the experiments with experiments derived by the use of fully supervised learning algorithm and the self-paced learning algorithm,it is shown that the self-paced learning with diversity can effectively improve the performance of the model.For the theoretical analysis,by decomposing and analyzing the loss function of self-paced learning with diversity,the effectiveness of the proposed algorithm is proved by considering the existing convex optimization and self-paced learning theories.Thirdly,we applied the proposed algorithm into three different medical image segmentation tasks and validated the approach on 5 datasets.In detail,the proposed method has been applied on retina vessel segmentation,lung organ segmentation and nuclear cell segmentation tasks.Experimental results indicate that the proposed method could significantly improve the performance of CNN based image segmentation model.By using the proposed method,the models can achieve a higher Dice score,surface distance and mean Intersection over Union on the same training dataset.