Research On Medical Image Semantic Segmentation Based On Semi-supervised Learning

Medical image semantic segmentation can accurately identify important organs and pathological tissues in images,thus it plays a significant role in computer-aided diagnosis,surgical planning,and intelligent healthcare.Recently,supervised learning-based image segmentation methods have achieved significant results.However,the success of these techniques critically depends on large amounts of accurately labeled data.In fact,in the field of medical image segmentation,the data resources for accurate labeling are very scarce.The reason is that the medical image annotation task requires not only a lot of professional knowledge,but also the image edges are blurred and noisy due to the differences in imaging.The annotation task requires a lot of manpower and material resources,and the cost is expensive.Compared to supervised learning,semi-supervised learning is a learning paradigm that addresses the problem of incomplete labeling in weakly supervised learning.It mainly utilizes a small amount of labeled data and a large amount of unlabeled data for joint learning,reducing the dependence of supervised learning networks on labeled data.semi-supervised learning mainly solves the problem of medical image segmentation under the constraint of labeled data resources,which is more in line with practical clinical application scenarios.Therefore,researching medical image semantic segmentation based on semi-supervised learning is of great significance.This thesis focuses on studying semi-supervised medical image semantic segmentation methods based on deep learning,aiming to solve problems such as inadequate use of unlabeled data,slow feature learning,low segmentation accuracy,and large model parameter sizes.The main work are as follows:(1)To addresses the issue of poor segmentation results due to the inability of conventional encoder-decoder segmentation networks to extract global contextual information and low efficiency of feature fusion,a semi-supervised medical image segmentation method based on cross self-attention and feature pyramid(FPS-Net)is proposed.First,a feature pyramid module(FPM)is designed in the encoding stage,which obtains the image context information through pyramidal multi-scale group dilated convolution,effectively solving the problem of variable shapes and sizes of organs and lesion regions in medical images.In addition,the pyramidal grouped convolution can effectively improve the network’s inference speed.Second,a cross-self-attention module(CSA)is designed in the feature fusion stage.This module firstly calculates the self-attention weights at the decoding stage,and then applies the weights to the encoding stage,and finally provides a finer attention feature map for the decoding stage by skip connection.The CSA module realizes the interaction of visual information between the spatial and channel of the encoder and decoder and obtains the remote dependencies of the images.Experiments show that the proposed FPS-Net can not only effectively improve the segmentation accuracy,but also enhance the inference speed of the network.(2)To address the problem that the mainstream semi-supervised medical image segmentation methods ignore the a priori relationship between labeled and unlabeled data,resulting in poor quality of pseudo labels and large number of network parameters,this thesis designs a semi-supervised medical image segmentation method based on adversarial consistency learning and dynamic convolution network(ASE-Net).First,a novel adversarial consistency training strategy(ACTS)is designed,which proposes a multi-task dual discriminator based on consistency and adversarial learning.The first discriminator mainly obtains the a priori relationship between labeled and unlabeled data.The second discriminator computes the image-level consistency of the unlabeled data under different data perturbations.The strategy effectively improves the utilization efficiency of unlabeled data and the prediction quality of pseudo labels mainly through pixel-level consistency and image-level consistency.Second,a dynamic convolution-based bidirectional attention component(DyBAC)is designed,which dynamically adjusts the parameters of the convolution kernel using information about the input samples’ own structure.The component achieves dynamic changes of network parameters,which can be embedded in any segmentation network to fully exploit the prior knowledge of the samples.Experiments show that the proposed ASE-Net is a lightweight network structure compared to state-of-the-art methods,which not only achieves high-quality segmentation results with improved feature representation but also a significant reduction in the number of parameters and computational complexity.