Research And Application Of Semi-Supervised Segmentation Method In Medical Image Assisted Therapy

Image segmentation technology plays a crucial role in many fields,affecting the accuracy of downstream tasks.With the development of deep learning in medical imaging,image segmentation technology has gradually become the essential means of adjuvant therapy.The method based on image segmentation technology can quickly identify the lesion area in medical images and assist doctors in diagnosis and treatment.However,the image segmentation technology based on supervised learning requires a large amount of precisely labeled data sets,and such data sets require experienced doctors to label,which makes it very difficult to collect accurately labeled data sets,so the effect of image segmentation and accurate rates are facing enormous challenges.Image segmentation techniques based on semi-supervised learning are more widely used to alleviate the problems posed by the dataset.This thesis mainly studies the application of image segmentation technology based on semi-supervised learning in medical assisted treatment tasks.It mainly expands the consistency regularization technology and self-training method in semi-supervised learning.Based on the characteristics of diseases in different parts of the body,the design is suitable for different disease or organ models,thereby improving the accuracy of segmentation and helping to perform downstream adjuvant treatment tasks.The main research content of this thesis is as follows:(1)This thesis proposes a new semi-supervised learning method based on consistency regularization,the multi-scale consistency self-training semi-supervised orbital tumor segmentation framework MSCINet.For the first time,semi-supervised learning is applied to segment orbital tumor lesions.This method can adapt to the characteristics of orbital tumor lesions with different sizes and significant changes in shape and has better generalization.The performance of the two data sets surpassed the previous method:the Dice index on the orbital tumor binary segmentation data set(Orbtum-B)data set reached 74.60%,exceeding 4.35%of the baseline method;in the orbital tumor and surrounding The tissue segmentation data set(Orbtum-M)also shows good performance,and the Dice value can reach 75.04%;(2)This paper proposes a semi-supervised segmentation framework MVCLGNet based on multi-view multi-scale local-global consistency,mainly applied to the semi-supervised segmentation of human abdominal gastrointestinal organs.It proposes a multi-view fusion module introducing two additional view images to mine richer semantic information.It addressed a loss in segmentation performance for gut organs due to positions stacked on each other.It has shown good performance on the MV-UWM dataset of the human abdomen and stomach segmentation.It has at least a 3.17%improvement in Dice value compared with the previous semi-supervised segmentation network;(3)This thesis proposes a system for predicting the curative effect of neoadjuvant chemotherapy for gastric cancer and constructs a curative effect prediction model RP-Model based on multi-omics fusion.The omics model showed excellent performance.The model achieved AUC above 80%on both the test and external validation sets.This prediction system has successfully moved the prediction window forward,which can predict the effect of neoadjuvant chemotherapy on patients,which is of great significance in the adjuvant treatment of gastric cancer.