CT Projection-Image Dual-Domain Features Based Automatical Bodypart Recognition

In clinical practice,many automatic medical image analysis algorithms need to be designed with prior knowledge of specific anatomical features.However,different organs usually have different anatomical features.Therefore,before performing many automatic medical image analysis algorithms,it is usually necessary to automatically recognize the bodypart information contained in the medical image to obtain prior knowledge.In addition,given the bodypart information,we can ofen reduce the search range for both detection and segmentation,and thereby improving the speed and robustness of the medical analysis algorithms.The technique of automatically recognizing the position of organs is known as bodypart recognition(BPR),which accurately and automatically locates the body position of a specific organ.Bodypart recognition is a prerequisite for the success of many medical image analysis tasks and computer-aided diagnosis(CAD)systems.However,the task of automatical bodypart recognition is challenging,and further research is still needed.Computed tomography(CT)is one of the most commonly used clinical imaging modalities.For a long time,researchers have proposed many bodypart recognition methods based on CT images,and achieved promising performances.As known,CT images are obtained by reconstructing the CT raw data(CT sinogram).Currently,studies of bodypart recognition based on CT sinogram are limited.However,the recognition of bodyparts based on CT sinogram is also of great significance.Because some effecitve feature information for bodypart recognition might be lost in the CT images due to the reconstruction process,and this part of the information might still exist in the CT sinogram.However,given that CT sinogram is difficult to describe and interpret for human experts,it is hard to use the traditional bodypart recognition algorithms to extract the corresponding features.In recent years,since the deep learning(DL)techniques,which can effiectively learn usefull features from target data,has been widely used in medical image analysis tasks,bodypart recognition based on CT sinogram is also becoming feasible.Therefore,in this article,we have conducted two researches exploring on how to effectively incorporate the CT sinogram information to improve the performace of bodypart recognition.(1)This article used convolutional neural network(CNN)to discriminate the five most common bodyparts(i.e.,head,neck,chest,upper abdomen,and pelvis)based on CT sinograms and CT images,and studied the recognition performance of CT sinogram.The CNN classifier with CT sinogram as input is denoted as Sino-Net,and the CNN classifier with CT image as input is denoted as Img-Net.The experimental results show that the performance of CT sinogram is similar to that of CT images in bodypart recognition and even better than CT images under some low-dose scanning conditions.(2)This article studied whether the feature information extracted from both CT sinogram and CT image(dual domain)can be complementary,and studied the recognition performance based on fusion features of dual domain under different scanning protocols.Specifically,we proposed a multi-class fusion algorithm based on evidence reasoning with reliability to fuse the output features of Sino-Net and Img-Net.The experimental results show that the dual-domain information is complementary,and the proposed fusion algorithm can effectively fuse the feature information of both CT sinogram and CT images and further improve the performance of bodyparts recognition.