Lymphoma Segmentation For 3D PET Images Based On CNNs

Position Emission Tomography(PET)is a main diagnostic tool for lymphoma.Due to the fact that the shapes,sizes and locations of the lymphoma are not fixed,and the involved lymphoma may be found throughout the body,in addition to those,the low resolution,high noise and fuzziness feature of the PET image and other factors,PET image-based lymphoma segmentation is a very challenging task.In recent years,deep learning technology has made breakthrough progress in the field of medical image segmentation.In view of this,this thesis proposes a series of data enhancement methods for 3D PET images based on the characteristics of lymphoma data to segment lymphoma based on Convolutional Neural Network(Convolutional Neural Network,CNN).The results show that the method proposed in this thesis effectively improves the segmentation performance,especially in the local dense region,the main work is as follows:For the problem of inaccurate target feature extraction caused by small data samples,this thesis explores the feature attention transfer of CNNs to improve the accuracy of target feature activation.Through the experimental analysis of Natural image,a data processing method for feature attention transfer is proposed.The foreground features are enhanced by using the cross-background method.The CAM(class activation map)shows that this method can effectively transfer the feature activation area to the target area.Finally,the method was verified in the segmentation of lymphoma,and its effectiveness was verified through a series of comparative experiments.Secondly,in view of the serious problem of missed segmentation of dense lesions,this thesis analyzes the reason from the perspective of CNN's sensitivity to local regions,and proposes a non-self-copy data enhancement strategy.The basic idea is to firstly build a lymphatic cancer library containing all the lesions in the training set,and then copy randomly the lymphatic lesions from the library to the background of the PET image during each iteration training,which makes new PET image data form dense lesions to improve the model's ability to segment dense lesions.The density can be controlled by the number of randomly selected lymphomas.After a series of experimental tests,this method can improve the performance of different network models such as U-Net,U-Net based on Residual Block and U-Net based on Dense Block,and effectively reduce the phenomenon of missed segmentation.Thirdly,in view of the problem of under-segmentation at the edge of the local lesions in the initial segmentation results,this thesis proposes a new method of fine segmentation.The dynamic ROI threshold method is used to optimize the initial segmentation results of single lesions in dense areas.Optimization is divided into two key points.First,the method of double prediction accuracy is used to ensure the accuracy of ROI selection,and to ensure that the selected area contains the entire lymphoma.Secondly,for the single lymphoma lesion newly appeared in the optimization process,it is regarded as a new ROI area,so that to adjust the fine segmentation of the edge of single lymphoma in dense areas.