Spinal Canal Segmentation,Reconstruction And Disease Diagnosis Based On Deep Learning

In recent years,spinal canal stenosis has become one of the most common diseases in clinical orthopedics.The main lesion of spinal canal stenosis is morphological changes,with the reduction of effective volume of spinal canal,causing compression of the nerve root,resulting in patients feel weakness or numbness in legs,calves and buttocks,giving great pain to them.Even sometimes,spinal stenosis can cause patients to lose control of some of their body functions.With continuous improvement of medical imaging technology,computed tomography(CT)is widely used in diagnosis of spinal stenosis.In clinical practice,it is usually necessary for a clinician with rich theoretical knowledge and practical experience to manually segment the spinal canal region from CT image of the patient and measure the anterior-posterior diameter of spinal canal to diagnose whether the patient has spinal canal stenosis.This process is very time-consuming and exhausting for clinician,and affected by accumulated knowledge,practical experience and subjective factors from different doctors,the diagnosis will have a larger error,need to repeatedly check to confirm.Therefore,in order to achieve full-automatic,rapid and accurate image positioning,segmentation and measurement of spinal canal,obtain more accurate diagnosis result,and also in order to reduce the work burden of doctors,save unnecessary manpower and time consumption,improve work efficiency,it is necessary to use computer technology assists doctors in performing preliminary disease diagnosis.Traditional medical image segmentation method,based on image processing,mainly use the shallow features of image such as grayscale,texture and statistical structure for segmentation.These shallow features are not representative,robust and easy to be disturbed by noise.With the continuous development of deep learning technology in recent years,great breakthroughs have been made in its application in the field of image.Many well-designed neural network structures have emerged in image recognition,target detection and image segmentation.Practice has proved that deep learning technology can mine the deep abstract features of data,which can improve the segmentation precision and achieve rapid and accurate automatic segmentation.At present,many applications of deep learning in images are based on a large number of training samples.For example,Image Net has more than 15 million artificially-tagged pictures and more than 22,000 categories.Based on this,the neural networks trained in image classification,object detection,object recognition accuracy are far more than human.However,in the field of medical imaging,it is often impossible to obtain such a large number of manual annotation data,which also brings new challenges to the application of deep learning in medical image field:1.Traditional spinal canal image segmentation methods are mostly semi-automatic methods and require a lot of manual intervention,poor efficiency in batch operation,and cannot make fast diagnosis.The method based on statistics and geometric structure cannot describe the morphological changes of the spinal canal region in the continuous CT scan of the spine.2.Deep neural network training requires massive data as a training set,but the medical image field has not yet formed a unified image annotation specification and cannot provide massive training data,resulting in insufficient training samples for neural network.3.The diagnosis of spinal stenosis disease requires doctors to have rich practical experience,but the medical treatment level and the experience and technical level of doctors in different regions are uneven,therefore,the measurement error is relatively large,and the accuracy of measurement and disease diagnosis cannot be guaranteed.At the same time,it requires doctors to do a lot of repetitive work,which seriously consumes the doctors' time and energy.To solve the above problems,this paper proposed a deep learning method for spinalcanal segmentation,reconstruction and disease diagnosis.We combine the classic convolutional neural network Faster-RCNN in target detection with U-Net,a fully convolutional network commonly used in medical image segmentation,and improve the U-Net network to make it a Shallow U-Net,which not only solves the central spinal canal in the position of lumbar spine CT image uncertainty,but also overcomes the lack of a large number of training data in the field of medical image,effective use of a small amount of sample data training model to obtain accurate image segmentation results.As the spinal canal image after segmentation is a local area in lumbar spine CT,and not completely left-right symmetry,that's why we use the principal component analysis to perform image correction on the segmentation result,then measure the rotational spinal canal to get the anterior-posterior diameter length,finally perform three-dimensional reconstruction of all segmentation results.The measurement results are mapped directly to the three-dimensional model based on the determination of the stenosis of the spinal canal,to assist clinicians in preliminary diagnosis of the disease,saving a lot of manpower and time,improve disease diagnosis effectiveness.After experiments,the deep learning network proposed in this paper can accurately locate and segment the spinal canal in CT images.The average Io U(Intersection over Union)reaches 88% and the average Dice reaches 93.6% in experimental data.In the measurement of spinal indicators,compared with the clinician's manual measurements,the anterior-posterior diameter measurement error is 0.57 mm and the right-and-left diameter error is 1.58 mm on average.The measurement error of anterior-posterior diameter which plays a decisive role in disease diagnosis process within clinical permitted range,preliminary diagnosis of spinal stenosis can be performed.