Algorithms For Localization And Segmentation Of Intervertebral Discs In MR Lumbar Spine Images

In recent years,with the development of mobile phones,computers and other technologies,people are sitting for a long time in life.The consequence of this is that lumbar spine diseases are highly prevalent in the population,becoming one of the main diseases affecting people’s daily life.When diagnosing lumbar spine diseases,it is often necessary for professionals to manually mark the position of the intervertebral disc.This method is outdated today when image data is increasing day by day.This paper studies the algorithm of localization and segmentation of intervertebral discs on T2-weighted Nuclear Magnetic Resonance lumbar spine images with less manual interaction or even a fully automatic method,which shortens the time of manual annotation and reduces the workload of doctors.The research contents of this article are given follows:(1)In the process of intervertebral disc positioning for MR lumbar vertebra images,an intervertebral disc localization algorithm based on broad learning system is proposed for the characteristics of uneven gray distribution of the image and complex background.After training the network model,we use a sliding window to traverse the images to be located.The traversal principle follows from left to right and from top to bottom.Then determine whether the image block in the sliding window is an intervertebral disc through the network.Finally,the model records the positioning coordinates of the intervertebral disc.When determining the final positioning coordinates of the intervertebral disc,linear regression removes the wrongly isolated coordinates.Then use fuzzy C-means clustering algorithm to cluster the positioning coordinates.According to the statistical characteristics of the intervertebral disc space in the lumbar spine image,the positioning coordinates of the first intervertebral disc are adjusted to obtain the final intervertebral disc positioning point.(2)When segmenting the MR lumbar vertebral discs,a CV model with a priori shape was proposed for the characteristics of partial intervertebral disc adhesion to soft tissue,blurred left edges,and blurred upper and lower edges.This method introduces the prior shape into the energy functional of the classic CV model,and constrains the evolution of the curve with the prior shape to improve the segmentation accuracy.In extracting the prior shape,a method combining outer contour and nucleus pulposus was proposed.According to the characteristics of extra lines on the left edge of the intervertebral disc,a convolution operator was used to remove adhesions,and the OTSU algorithm was used to calculate the threshold value.Threshold segmentation was performed on the image to obtain the outer disc segmentation result.Aiming at the characteristics of blurring of the upper and lower discs of the disc in the image,the difference between the processed image and the original image is used to enhance the edge of the image.Then the morphological operation and region growth are used to obtain the nucleus pulposus segmentation result.Extracting the prior shape of the intervertebral disc is superimpose the obtained outer contour and the segmentation results of the nucleus pulposus,and then fill it with morphology.(3)For the entire MR lumbar image,this paper uses an improved U-Net network to segment it.Because the U-Net network layer is not deep,there are a lot of over-segmentation and under-segmentation in the segmentation results,and the segmentation accuracy is not high.This paper improves the segmentation effect by increasing the number of network layers.In order to avoid problems such as network performance degradation and excessive number of parameters,the concept of "shortcut"in Residual Network and a new shrinking network is introduced into U-Net network,to build a new network model ResU2/3-Net.Take the complete MR image and the corresponding gold standard as the training set input,train the network model,and output the predicted intervertebral disc region probability map.This method realizes fully automatic disc segmentation and improves segmentation accuracy.Experimental data shows that on the MR lumbar spine image,the intervertebral disc location algorithm based on broad learning system can effectively locate the disc area.The average Euclidean distance between the manually marked positioning point and the positioning point obtained by this algorithm is 3.9865,which avoids wrong positioning Leak positioning.Both the shape-a priori CV model and the improved U-Net network can obtain more accurate intervertebral disc regions.The DICE coefficients are 0.9225 and 0.9357,respectively,and the segmentation accuracy rates are 96.38%and 95.73%,respectively.