Research On 3D Medical Image Segmentation Algorithm Based On Semi-Supervised Learning

Accurately segmenting tissue structures or pathological regions in medical images plays a crucial role in the diagnosis and treatment planning by physicians.However,fully supervised learning algorithms rely on a large amount of annotated data for training,and medical images are highly complex and diverse,making accurate annotations time-consuming and labor-intensive.This severely limits the application and development of fully supervised learning in the field of medical image segmentation due to the limited availability of labeled data for training.This thesis aims to study a semi-supervised 3D medical image segmentation algorithm under the constraint of limited annotated data.It primarily addresses the issues and challenges in medical image segmentation tasks.Firstly,supervised learning algorithms tend to have lower segmentation accuracy on medical images with scarce annotated data.Secondly,the surfaces of segmented organs or tissues obtained from semi-supervised methods lack smoothness.Lastly,existing semi-supervised segmentation algorithms may exhibit uncertain prediction in non-target regions,resulting in the inclusion of irrelevant and trivial parts in the segmentation results.To address these issues,the main research content and innovations of this thesis include the following three points:(1)A semi-supervised 3D medical image segmentation framework based on feature enhancement and interactive learning is proposed to address the problem of sparse of medical image annotated data.The framework embeds a feature attention module into the segmentation network to help the model extract highly discriminative features.Based on the average teacher model,a dual-task network is utilized to predict both pixel-level segmentation map and Symbolic Distance Map(SDM).The SDM serves as a common constraint for annotated and unannotated data,and semi-supervised learning is achieved through consistency regularization to accomplish high-precision segmentation.The framework achieved good segmentation results on the left atrium(LA)dataset.(2)A semi-supervised 3D medical image segmentation method based on boundary geometry constraint is proposed to improve the segmentation accuracy of boundary regions and to make the segmented organs or tissues have a smooth surface.This method based on a dual-task network embeds a reverse attention module into the V-Net decoder,and can simultaneously predict the pixel-level segmentation map and background map of the target object.For unlabeled data,the contour information of the target object is obtained by element-wise multiplication of its segmentation map and background map,and geometric constraints are applied to semi-supervised learning using this contour information.Experimental results on the LA,pancreas(NIH Pancreas),and brain tumor(Bra TS2019)datasets demonstrate that this algorithm can effectively capture the geometric shape of the target object and ensure the smoothness of the segmented result boundaries.(3)A semi-supervised 3D medical image segmentation algorithm based on pseudo-label uncertainty suppression is proposed to reduce uncertainty prediction during the segmentation process,reduce irrelevant and trivial parts in the segmentation results,and further improve the segmentation accuracy.The algorithm obtains uncertain and certain regions by performing a logical XOR operation on the segmentation map and the background map,which are used as pseudo-labels to calculate unsupervised and consistency losses for uncertainty suppression.The weighted cross-entropy loss function is used to calculate supervised losses,assigning higher weights to the boundary regions to enhance the model’s perception of the boundaries and further reduce the uncertainty in these regions.The pseudo-label data is updated in a cyclic iterative method to improve the quality of the pseudo-labels and guide the model to perform semi-supervised learning better.Experimental results on the LA,NIH Pancreas,and Bra TS2019 datasets show that the proposed algorithm can significantly improve the accuracy of medical image segmentation.