Research On Spinal Localization Method Based On Deep Learning And Spatial Relation Reasoning

With the progress of medical visualization,the development of medical images has brought great convenience to clinical diagnosis,and the application of computer vision in medical images has been more and more extensive,which has important research significance and value.At present,spinal diseases are becoming more and more widespread in the population,and mri has a better display of clinical diagnosis than other forms of imaging.The positioning and labeling of intervertebral disc is the focus of current research.The traditional manual positioning and labeling method is time-consuming and has a certain error rate,so it is of great significance to use computer for auxiliary diagnosis.The current methods of positioning and labeling related to spinal image have some shortcomings.They are all trained for specific morphology,image sequence or parameter setting,and they are not universal.Therefore,based on deep learning and spatial reasoning,this paper studied the new method of disc location labeling,so as to realize the location and labeling of disc in any position,as well as local or global images.First,RefineDet target detection algorithm was used to locate the spinal MRI image,obtain the initial position of the disc,introduce advanced anatomical constraints of the disc,screen the initial position,and obtain an accurate collection of disc positions.Secondly,according to the spatial reasoning knowledge,every three adjacent intervertebral discs were modeled,and 21 adjacent intervertebral disc spatial relationship models were trained.Finally,according to the precise disc location set and the spatial relationship model of adjacent discs,the iterative reasoning algorithm of disc spatial relationship was proposed and the iterative matching was carried out.The main research work of this paper is as follows:1.Initial positioning of the discThe deep learning method is used to extract the underlying features of the image and identify the target,and all the candidate positions of the disc are obtained to form a corresponding candidate set.A scoring constraint algorithm is proposed to screen the candidate sets and eliminate the error detection,so as to obtain the precise set of disc positions.2.Proposed the spatial relationship model of adjacent intervertebral discsBased on the geometric structure and anatomical structure of the human spine,the appearance and local spatial relationship of the three adjacent discs were modeled.The model features mainly include two aspects: image intensity characteristics and morphometric characteristics.A total of 21 adjacent disc spatial relationship models were trained.3.Proposed the intervertebral disc spatial relation iterative reasoning algorithmBased on the precise location set of intervertebral discs and the spatial relationship model of 21 adjacent intervertebral discs an iterative reasoning algorithm for intervertebral disc spatial relationship is proposed.Multiple initializing marker sequences were obtained by initializing the collection of intervertebral discs,and the matching degree of each sequence was obtained by iteratively matching with the corresponding model from the first node of the marker sequence.Finally,the sequence with the best matching degree was selected to label the disc.