Algorithm Design And Application Research For Deep Semi Supervised Medical Image Semantic Segmentation

With the rapid development of computers in the medical field,techniques for identifying and segmenting lesion regions from medical images have been extensively studied and applied,and achieving computer-aided diagnosis is of great importance in clinical practice.In recent years,medical image recognition methods based on deep learning have become a popular research topic.However,to obtain high-performance network models in the medical imaging domain,it relies heavily on a large amount of labeled data.In particular,in medical image segmentation tasks,labeled data requires physicians with strong expertise to perform accurate pixel-level annotations.Data labeling is very expensive and time-consuming,so the amount of labeled data in medical images is usually small,which leads to a significant reduction in the performance of supervised learningbased models.In this case,efficient use of unlabeled data becomes crucial in the medical image segmentation domain.In this paper,we investigate the problem of medical image segmentation with a small amount of labeled data and a large amount of unlabeled data using semisupervised learning methods and propose two deep semi-supervised methods for medical image segmentation.The specific work is as follows:1.To improve the model’s utilization of unlabeled data,a self-training based consistent semi-supervised medical image segmentation algorithm is proposed.The proposed model is a dual-teacher model augmented with strong and weak data augmentation and learns unlabeled data features through a dual-teacher consistency loss.The strong and weak teacher model is used to generate pseudolabels to expand the labeled data set,and the self-training method is used to increase the model’s recurrent learning of unlabeled data.Better segmentation results have been achieved on the skin lesion image dataset(ISIC),the retinal fundus image dataset(Refuge),and the low-grade glioma image dataset(LGG).2.Aiming at the problem of inductive bias in the mean teacher model,a semisupervised medical image segmentation algorithm based on consistency regularization and the adversarial mechanism is designed.This algorithm introduces an adversarial mechanism based on consistency regularization and generates unlabeled data through a discriminator.The confidence map provides more supervised information for the model.At the same time,the collaborative learning of daul discriminators is used to introduce unlabeled data into the training of discriminators to help the adversarial mechanism generate high-quality supervision information.The experimental results show that the method based on consistency regularization and adversarial mechanism has achieved better performance.3.Based on the above method,a deep semi-supervised medical image intelligent segmentation system is designed,and the segmentation model is packaged into the terminal device to realize real-time intelligent segmentation of lesion areas to achieve a computer-aided diagnosis...