Research On Automatic Diagnosis Of Lumbar Disc Herniation Based On Deep Learning

Nowadays,lumbar disc herniation is one of the most important health problems in social medicine and a common cause of mobility problems in middle-aged and elderly people.With the increase of work pressure and the acceleration of social rhythm,the number of "sedentary people" is increasing,the number of people suffering from lumbar spine related diseases is increasing one after another,and the trend of lumbar disc herniation is also showing a younger age.The treatment methods for lumbar spine diseases are becoming more and more refined,and the clinical requirements for the diagnosis of herniated discs are getting higher and higher,and the research on computer-aided diagnosis is particularly urgent.Therefore,in this paper,computer-aided technology based on deep learning is used to study lumbar spine MR images,and the whole process management from classification and screening of the original image dataset,image segmentation and diagnosis of herniated symptoms is realized,and the main research contents are as follows.Firstly,for the classification and screening of the original image dataset,this paper proposes a model combining SVM with SSIM and ORB,respectively.In the first step,by analyzing the characteristics of the original MR image dataset,SVM is used to initially classify the different signal and orientation images in the original dataset;in the second step,different image similarity matching methods are used to further screen the initially classified image dataset,and finally,the combination model of SVM and SSIM algorithm is used to classify the sagittal image dataset,and the combination model of SVM and ORB algorithm is used to The model was used to classify the cross-sectional image dataset.Secondly,to achieve semantic segmentation of lumbar vertebrae and intervertebral discs,a segmentation method based on improved Attention-U-Net is proposed in this paper.The model uses an attention module based on multi-level feature fusion and introduces two residual modules instead of the original convolutional blocks.After segmenting the target region,the vertebrae and intervertebral discs are extracted using gray-scale thresholding,and image fusion is used to achieve image overlay.Then,an improved PSO-SVM classification algorithm and LDH-YOLO model are proposed for the classification of lumbar disc herniation symptoms to provide the aid of LDH diagnosis for medical patients.The former mainly improves the classification accuracy of SVM by determining the optimal parameters of SVM using particle swarm algorithm(PSO).The latter improves the structure by analyzing the composition structure of different versions of YOLO and combining its characteristics with the specific conditions of the lumbar spine,while adding Mosaic data enhancement and Label Smoothing to prevent overfitting in the training method,and introducing CIOU and learning rate cosine annealing decay to make the training process more stable.The LDH-YOLO model is proved to have more stable recognition effect through comparison experiments,and the F1-Score reaches more than 83%.Finally,the lumbar spine diagnosis system is designed to realize the whole process management of lumbar spine diagnosis.The workflow of the system is sequential around the order of full process management,including three modules: first,classification and screening of clear images;second,semantic segmentation of lumbar vertebrae and intervertebral discs;and third,implementation of the algorithm model for LDH classification.The system interface is oriented to orthopedic surgeons and consultation personnel,so the interface has the characteristics of simplicity and intuitiveness,which facilitates the user to use the system conveniently.