Boundary Focused Salient Object Detection And Its Application In Medical Image Segmentation

Both salient object detection and medical image segmentation are important sub-tasks of Image Segmentation.The goal of image salient object detection is to locate the most attractive area in an image.As a preprocessing step of major computer vision tasks,image salient object detection has attracted more and more attention in recent years.Firstly,different scale features contain different information.Appropriate multi-scale feature fusion is of great significance for salient object detection.However,most methods use direct concatenation or addition operation to process multi-scale features,which ignores the diversities of the contribution of different scale features to the generation of the final saliency map.Secondly,because the size of salient objects in the image is different and most salient object detection models can not dynamically adjust the receptive field to adapt to objects of different sizes,which leads to insensitivity to the size of salient objects.Finally,previous work mainly focused on feature extraction from the global view rather than distinguishing between local or pixel features,that is,most of them focused on the accuracy of the salient object region and ignored the quality of the boundary.The main goal of medical image segmentation is to identify the pixels of organs or lesions from the medical image and classify different regions of the image into predetermined categories according to the features extracted.It is one of the most challenging tasks in medical image analysis.At present,the main trend of medical image segmentation is to improve the U-Net network architecture.However,continuous convolution will inevitably lead to the loss of relevant information,which is similar to the problems of salient object detection in the field of natural image segmentation,that is,the segmentation of medical organs or lesion areas to be segmented is incomplete,and the detail processing is not perfect.To address the above problems,we made the following two aspects of work:(1)In order to solve the related problems in the field of salient object detection,a Boundary Focused Progressive Selection Network(PSNet)is proposed in this paper.Firstly,in order to dynamically select appropriate receptive fields to adapt to salient objects of different sizes,a Pyramid Feature Dynamic Extraction module is proposed to extract the high-level semantic features of salient objects;Secondly,in order to consider the boundary quality of salient objects from both global and local perspectives,a Self-Interactive Attention module is designed to extract the details of low-level edge features;Finally,a Scale Aware Fusion module is proposed to refine the edges of salient objects by making full use of the extracted advanced features.The experimental results show that compared with the salient object detection methods in recent years,the proposed method has good performance in both qualitative and quantitative experiments.(2)Aiming at the problems of incomplete segmentation of medical organs or lesions and imperfect detail processing in medical image segmentation,this paper applies the method of salient object detection to the field of medical image segmentation and modifies the deep learning algorithm according to the characteristics of medical image.Extensive experiments show that the improved Progressive Selection Network proposed achieves excellent results in medical image segmentation.