Research And Application Of Semi-Supervised Segmentation Algorithms For MRI Medical Images

Image segmentation is one of the fundamental tasks in medical image analysis.In the field of image segmentation,fully supervised learning methods have been widely applied.They can accurately segment important structures such as organs and lesion regions from medical images,which greatly assist doctors in diagnosis.Moreover,precise segmentation results have a positive impact and provide guidance for subsequent tasks.However,fully supervised learning methods require a large amount of annotated data to train the models and achieve good performance.In the field of medical imaging,acquiring annotated data is costly due to the need for specialized medical knowledge and skills.Furthermore,the quality of annotation is difficult to guarantee.Additionally,medical image datasets are typically small,and they suffer from issues such as image non-uniformity and class imbalance,which also affect the performance of fully supervised learning methods.Finding ways to learn and achieve good results from limited annotations has become a hot topic in the field of medical computing.In response to this challenge,semi-supervised learning methods have emerged as effective solutions.These methods can effectively learn from limited annotations and achieve algorithm performance comparable to fully supervised learning methods,significantly reducing the workload of annotators.Through in-depth research,this paper proposes improvements to existing semi-supervised medical image segmentation algorithms by leveraging unlabeled data for model training,aiming to address the challenges posed by the lack of annotated information and uneven image quality in medical image datasets.The main research work and contributions of this paper can be summarized as follows::(1)This paper proposes a semi-supervised medical image segmentation model called MSCNet(Multi-Scale Consistency U-Net),based on multi-scale consistency regularization.The model draws inspiration from the core ideas of the CCT(Cross Consistency Training)semi-supervised method and applies the concept of auxiliary decoders to the different scales in the U-Net backbone network,generating multi-scale prediction results to leverage multi-scale information for semi-supervised segmentation.To enhance prediction stability and reliability,this paper introduces a multi-scale consistency loss to enable mutual supervision among the auxiliary decoders.Additionally,to enhance the performance and representational capacity of the model,this study improved the encoder architecture of the network by adopting the more effective Res Net structure.Furthermore,considering the class-imbalance in the dataset,this paper employs a mixed loss strategy for training the labeled samples.(2)To overcome the limitations of MSCNet in 3D medical image segmentation,a semi-supervised segmentation model called CAMT(Cut Mix Attention Mean Teacher)is proposed specifically designed for 3D medical images.The model utilizes a 3D U-Net architecture with a fusion attention module as the backbone network.The attention module consists of parallel channel attention and spatial attention modules connected through skip connections,enabling accurate feature capturing and enhancing segmentation accuracy.To fully leverage the unlabeled data,the Mean Teacher semi-supervised model is employed,along with a confidence-based selection module for training.Additionally,to increase the diversity of training data and improve the effectiveness of consistency regularization,the CutMix data augmentation method is applied to perturb the unlabeled data.(3)We designed and developed a medical image automatic segmentation system that encompasses the essential functionalities required by medical professionals and related personnel.It provides a user-friendly interface and offers excellent postexpansion capabilities.Additionally,the system integrates two segmentation algorithms proposed in this article.