Research On Medical Image Segmentation Methods Based On Multi-Expert Annotation

Medical image segmentation is a crucial field in medical image analysis and an indispensable tool for assisting diagnosis,detection,and treatment.In recent years,deep learning techniques have made significant advancements in computer vision and have been widely applied in the field of medical image segmentation.Automated medical image segmentation systems provide reliable references for disease diagnosis and offer detailed image information to aid doctors in making more informed diagnostic and treatment decisions.Due to the richness of information,fuzzy boundaries,and ambiguous regions in medical images,it is common to rely on multiple annotators for independent labeling to mitigate errors caused by varying levels of expertise or subjective biases.Studies in the clinical domain often report differences among annotators,which can pose challenges in segmenting highly uncertain regions.There is an increasing focus on the differences between expert annotations.In the field of computer vision,when dealing with multiple annotations,the commonly used methods such as majority voting or selecting the preferred annotator do not effectively capture the consistent and diverse information among multiple experts.Therefore,a new challenge arises: how to incorporate rich information from different expert annotations into the segmentation results.To address this issue,this paper investigates the information bottleneck method and variational autoencoder model,and extends them to the task of medical image segmentation with multiple expert annotations,aiming to extract both the consistent and diverse information between different views.Based on deep learning algorithms,this study conducts research on medical image segmentation with multiple expert annotations from multiple aspects including feature extraction and model performance.The main contributions of this study are as follows:(1)Design and Implementation of a Medical Image Segmentation Model Based on Information Bottleneck and Multiple Expert Annotations.To eliminate the noise within annotators and the redundant information among annotations,and to extract the common information among multiple expert annotations,this study proposes a medical image segmentation model based on multiple expert annotations.The information bottleneck method is theoretically extended to unsupervised multi-view information bottleneck,which maximizes the mutual information between multiple annotations to preserve consistent information and reduce redundancy.The multi-view information bottleneck method can be regarded as a novel strategy for multi-view representation fusion,focusing on extracting highly correlated features.To the best of our knowledge,this is the first model that extracts consistent information from multiple experts using the multi-view information bottleneck.The model is evaluated using sensitivity,specificity,accuracy,AUC,and Dice coefficient,among other evaluation metrics,and extensive experiments are conducted on both public and private datasets of medical images.The experimental results demonstrate that the proposed model outperforms existing techniques in terms of performance.(2)Design and Implementation of a Coarse-to-Fine Medical Image Segmentation Model Based on Variational Autoencoder.The difference information among multiple expert annotations also reflects important signals.To further explore the consistent and differential information among multiple expert annotations,enhance label reliability,and improve model generalization ability,this study proposes a coarse-to-fine medical image segmentation model based on the variational autoencoder.The model utilizes a probabilistic variational autoencoder to learn the prior and posterior features of the images.It employs a soft attention mechanism to focus on the uncertainty mapping between features and labels,integrating the consistency and diversity information among multiple expert annotations,thereby enabling the model to achieve more accurate representations.The coarse-to-fine medical image segmentation model based on the variational autoencoder provides an automated tool for medical image segmentation,facilitating the learning of annotation information from multiple experts and effectively improving the model’s robustness and generalization ability.In summary,this paper describes the significance and limitations of medical image segmentation and multiple expert annotations.It provides detailed explanations of the proposed medical image segmentation models based on information bottleneck and variational autoencoder to address these limitations.The research conducted extensive experimental evaluations,validating the effectiveness of the models.The experimental results demonstrated the feasibility of medical image segmentation techniques based on information bottleneck and variational autoencoder,laying a foundation for further research on label fusion strategies for multiple expert annotations.