Key Technology Research On Aided Diagnosis Of Spine Diseases Based On Deep Leaning

Currently,spinal deformity is a common disease that harms young people.The latest statistics shows that there are more than 3 million patients with spinal deformity in our country,and the incidence of spinal deformity in adolescents is up to twenty percents.The smaller the degree of deformity of the spine,the easier it is to be corrected,so early detection and treatment are crucial.The clinical diagnosis,follow-up,and surgical planning of spinal deformity rely significantly on the spinal alignment parameters,including coronal alignment(i.e.cobb angle)and the sagittal alignment(i.e.thoracic kyphosis,lumbar lordosis,and sacral slope).During actual diagnosis,the measurements of these parameters are commonly performed by professional doctors using the built-in tools of hospital-based picture archiving and communication system(PACS).The measurement is a manual and time consuming process.With the rapid increase in the number of patients’cases,the doctors’work pressure will increase sharply,that may increases the risk of inaccurate measurement.The measurement result presents an avoidable intra-and inter-observer variabilities.Thus,it is of great significance to study the high-accuracy and high-efficiency automatic spinal alignment analysis tools.In this thesis,the automatic measurement of the sagittal alignment uses the optical images of the mobile phone captured from actual clinical spinal X-ray images as the experimental datasets.The vertebral area of the thoracic spine in the sagittal position is affected by somethings,sucn as chest ribs and so on,which results that X-ray images of sagittal plane shows artifacts,and the shape of the vertebra is not obvious.In response to the above problems,the study in this thesis firstly locates the spine area and filters out the background area;and then,a segmentation network suitable for this study is designed to obtain images of vertebra by comparing multiple image segmentation networks;finally,considering the case of spinal deformity,the shape of the vertebra presents obvious irregular polygons.In order to solve this problem,an adaptive deep neural regression network based on ensemble learning is designed,so as to the measurement of the angles of sagittal spinal curve,thereby completing automatic measurement of the sagittal alignment.This thesis also studies the automatic measurement of the coronal alignment.Unlike the sagittal images,where each vertebra in the coronal X-ray image is clearly visible.According to the characteristics of these images,this study adopts the key point detection technology to detect the corners of the spine.Using the structure based on the network of VGG and ResNet as the image feature extractor,the study instroduces the heat map of the image,constructs the key point detection network modules,and designs the adaptive angle regression neural network,so as to completing the automatic measurement of the angles of coronal spinal curve.Through comparative experimental analysis and verification,you can find the above methods and modules make the automatic measurement and diagnosis of spine accurate and efficient,and achieve the expected results.