Medical Image Organ Segmentation Method Based On Position Prior

In various fields of medical research,such as radiotherapy target delineation,surgical planning,etc.,accurate organ segmentation is an indispensable and important part.With the development of deep learning,more and more medical organ segmentation methods based on convolutional neural networks have been proposed and achieved good results.However,most of the existing network methods have not explicitly modeled the local geometric information or position information of the segmentation targets(organs).Not only is it lack of interpretability and reliability,but it is also difficult to improve the failure mode of the network.Existing studies try to combine prior information to constrain the segmentation results of the network.Although the results have better interpretability,there are still some problems:most of these methods use bounding boxes to extract the region of interest and then refine the segmentation frame,but the boundary The use of boxes not only has positional deviation,but also cannot achieve end-to-end training;in addition,some studies constrain the network fitting results from the perspective of loss function,but the constraint ability of this method is very limited.In order to solve the above problems,this paper proposes an organ segmentation method based on prior information of position.The specific work is divided into two aspects:First,Ex Net method is proposed for the single organ segmentation,The organ contour extreme points(leftmost,top,bottom,and rightmost)and center points are used to model the appearance of the organs and establish a position prior model.In addition,a multi-task learning framework is used to simultaneously learn and predict the position information of the five points and the segmentation map information of the organ.Combining the position prior provided by the extreme points,the segmentation network can know the accurate segmentation position,thereby improving the reliability and accuracy of the positioning.Second,Re Net method is proposed for the multi-organ segmentation,The larger organs that are easier to be accurately segmented are used to guide the smaller organs with poor segmentation accuracy.Specifically,a two-stage network is used: in the first stage network,the relative positional relationship between large and small organs is modeled,and segmentation map information of large organs is first predicted.After that,the prior information on the large organs and their relative positions is used to determine the region of interest where the small organs are located and cropped from the original image.The second stage network is solely responsible for the refined segmentation of small organs,and the final results are merged.The experimental results show that the use of location prior information can effectively improve the segmentation accuracy of the network: in the challenge of CHAOS and Li TS two liver single organs,the method proposed in this paper obtained the segmentation accuracy of0.952 and 0.953,respectively;in Struct Seg2019,multiple organs In the segmentation challenge,the method proposed in this paper achieved a segmentation accuracy of 0.849,and the segmentation effect for small organs was greatly improved.The two methods proposed in this paper both model the location information and integrate it with the general segmentation network as the prior information of the constraint segmentation result,so the network has semantic reliability and robustness to a certain extent.