| In clinical medical practice,in order to formulate a treatment plan and accurately diagnose the condition,it is necessary to use medical imaging equipment to scan and examine the patient,and then submit the medical image to the doctor for diagnosis.This method relies heavily on human subjective judgment and requires doctors to have outstanding abilities and rich clinical experience.In recent years,with the rise of deep learning,the use of deep learning to segment lesions and assist in diagnosis has become an effective means to promote clinical medical analysis.Because medical images usually have the characteristics of blurred edges,uneven grayscale,noise and artifacts,etc.,it is still a difficult point to achieve high-accuracy lesion or organ segmentation.This paper hopes to explore the feasibility of improving the accuracy of the segmentation algorithm from the perspective of enhancing the learning of the edge region of the organ.In general,the main research and contributions of this article are as follows:First,a multi-input superpixel boundary assisted segmentation algorithm is proposed.A novel segmentation framework is proposed,which uses the appearance characteristics provided by superpixels to guide segmentation.The proposed segmentation framework extends the classic encoder-decoder segmentation network with additional bottom-up branches to extract effective information from superpixel boundary images;at the same time,a series of attention modules are proposed and studied to explore how the boundary information of superpixels is effectively merged into the main segmentation branch.Second,a super pixel boundary supervised learning algorithm based on multi-label learning is proposed.Firstly,integrate the atrous spatial pyramid pooling module into the encoder-decoder network structure as the basic framework;then the superpixel boundary and the original label are used as the supervision signal to constrain the network training at the same time,so as to mine the semantic information implicit in the superpixel boundary.That is,from the perspective of multi-label supervised learning,an algorithm based on superpixel boundary image supervision is proposed for medical image segmentation.Third,a two-stage segmentation method for medical images based on distance transformation is proposed.Using the edge of the left atrium as the target area for distance transformation,the obtained distance map can be used as a learning weight map to make the network pay more attention to the area near the edge of the organ.The training is divided into two stages in total.In the first stage,two branches are derived,and the rough left atrium segmentation and distance map prediction are performed respectively.The outputs of the two are merged into the second stage of training,and the accurate segmentation result is obtained in the second stage. |
