Research On Recognition Algorithm Of Lumbar Epidural Fat Based On Deep Learning

Lumbar pain is one of the diseases with the highest prevalence and disability rate in the world,and it is also one of the common reasons for orthopedic outpatients or inpatients.With the development of economy and medical technology,the rate of lumbar Magnetic Resonance Imaging(MRI)examination in patients with lumbar pain is gradually increasing.For doctors,the workload of MRI reading also increases correspondingly.However,due to the limitation of experience,speed and energy,the quality of film reading may decrease,and there is a risk of misdiagnosis or missed diagnosis.With the rise of artificial intelligence technology,detection methods based on deep learning have also been widely used in the medical field.Based on the classical deep learning object detection algorithm YOLOv4(You Only Look Once),this dissertation takes the epidural fat image in lumbar MRI as the research object,and focuses on the classification of epidural fat.The main research contents of this dissertation include:(1)In view of the selection of lumbar MRI epidural fat as the research object of this dissertation,a total of 2619 lumbar MRI images were collected.Firstly,they were divided into Data A and Data B.The Data A contained 1474 original images,and the lumbar epidural fat was marked.The data augmentation method was used to increase the number of images in the Data A to 6 times to 8844,as training validation and test set;Data B contains 1145 original images as a supplementary test set to verify the reliability and robustness of the algorithm.The constructed lumbar epidural fat data set was analyzed and sorted out,and multiple detection difficulties such as different shapes,different sizes,and unbalanced sample categories were obtained,which was convenient for subsequent targeted improvement of the network.(2)According to the characteristics of lumbar epidural fat dataset,a lumbar epidural fat recognition detection algorithm based on improved YOLOv4 is proposed.By analyzing a variety of object detection networks,CSPDarknet53(Cross Stage Partial Darknet53)is selected as the feature extraction network of this dissertation.The Atrous Spatial Pyramid Pooling is used to replace the Spatial Pyramid Pooling to avoid the decrease of feature layer resolution and improve the feature extraction ability.In order to solve the imbalance between positive and negative samples,difficult and easy sample classification,the Focal loss function is introduced and combined with the Binary Cross Entropy loss function to form a new loss function.The KMeans + + clustering algorithm is used to formulate the size of the anchor frame and optimize the anchor frame parameters of K-Means.The improved network model can achieve accuracy98.33 %,accuracy 96.36 %,sensitivity 94.34 %,and AUC(Area under the curve)value 0.974 in Data B.It has a good effect and provides a more accurate reference,which can assist clinical diagnosis.(3)Aiming at the problem of high complexity of network model,a lumbar epidural fat recognition detection algorithm based on improved YOLOv4-Tiny is proposed.CSPDarknet53-Tiny is used for feature extraction,a layer of shallow features is added,and a multi-feature fusion is used to construct a Feature Pyramid Network,which strengthens the utilization rate of network model feature information and solves the problem of missed detection.The Soft-NMS algorithm is used to replace the anchor box strategy of the original NMS algorithm.Instead of deleting the dense prediction boxes,the confidence is reduced to retain more prediction boxes,which improves the detection ability of overlapping adjacent prediction boxes to be detected and prevents missed detection and multiple detection.Aiming at the problem of artifacts in MRI,a Squeeze Excitation Networks and an Efficient Channel Attention module to YOLOv4-Tiny,to the detection of channel feature information of lumbar MRI.In the improved network model,Data B can achieve accuracy 97.59 %,accuracy 95.33 %,sensitivity 92.92%,AUC value 0.959.The experiments show that the improved network has improved in each evaluation index.