Research On Medical Image Segmentation Algorithms Based On Confidence Learning

Medical image segmentation has always been one of the important stages to assist doctors in clinical decision-making and improving diagnosis efficiency.Especially,the precise segmentation of tumors and other lesions can provide doctors with important evidence in clinical diagnosis and intraoperative resection.In actual clinical applications,doctors are required to annotate the medical image pixel by pixel to get the final segmentation result.However,manual segmentation of these medical images not only requires doctors with professional knowledge but also takes a lot of time(especially the annotation of 3D medical images).With the continuous development of deep learning and convolutional neural networks,the deep learning-based automatic segmentation model can greatly improve the efficiency of medical image segmentation through learning from large-scale labeled data.Although the automatic medical image segmentation algorithm based on deep learning can quickly generate image segmentation results,it still faces the problem of lack of high-quality data during the training process.Generally speaking,using a small amount of data for deep learning algorithm training often leads to overfitting of the model on the training data.Therefore,the training of deep learning algorithms requires a large amount of high-quality labeled data,and the quantity and quality of the data are usually one of the main bottlenecks that limit the ability of the model.However,there are a large number of unlabeled data in clinical practice.How to use these unlabeled data to assist model training is an important issue in current medical image segmentation tasks.In addition,some interactive medical image segmentation algorithms introduce the doctor’s error correction information on the basis of the original medical image segmentation algorithm,so that the model can correct the coarse segmentation results.This error correction information can also be regarded as an additional supervision signal,but this information is usually relatively sparse,and how to efficiently use it is also a difficult point in the current medical image segmentation research.To address the above challenges,the works of this article are as follows:1.To use these unlabeled data,this article proposes a semi-supervised medical image segmentation algorithm based on adversarial confidence learning,which utilizes the discriminant network in the framework of generating adversarial networks to learn the confidence of the segmentation results.The obtained confidence information can not only apply the adversarial loss in the training of the labeled data to regularize the segmentation network,but also calibrate the pseudo-labels of the unlabeled data,so that the model can reduce the negative impacts of pseudo-label bias on the model when using the pseudo-labels for training.2.To use the interactive information,this article proposes an interactive medical image segmentation algorithm based on confidence learning.To relieve the phenomenon that existing interactive segmentation algorithms are not sensitive to the error correction information during the interaction process,this article combines multi-agent reinforcement learning and confidence learning to evaluate the confidence of the segmentation action after the interaction,so that the proposed model can recognize those areas that are not sensitive to interactive information.3.To utilize feedback of error correction actions in interactive segmentation efficiently,this article designs a self-adaptive feedback mechanism,which uses the confidence evaluation results of error correction actions to weight the original reward functions in reinforcement learning,so that the model can pay more attention to those areas where the improvement is not obvious after the interaction during the training stage.Through this mechanism,the proposed model can more effectively use the evaluation results of the error correction actions,and continuously adjust the error correction strategy of the model during the interaction process.This article validates the proposed algorithms on multiple 3D medical image datasets.The experimental results show that the algorithms proposed in this article can not only surpass other medical image segmentation algorithms on different 3D medical image datasets but also can reduce dependence on labeled data,which proves the effectiveness of the proposed algorithms.