LSS-net: 3-dimensional segmentation of the spinal canal for the diagnosis of lumbar spinal stenosis

摘要

Lumbar Spinal Stenosis (LSS) is one of the main causes of chronic low backpain. Chronic low back pain not only reduces the quality of life of people butalso can be an important expense item in the country's economy due to theinability of the person to participate in working life and treatment costs. As inother diseases, rapid diagnosis and early treatment of LSS significantly affectthe quality of life of the person. Magnetic Resonance (MR) imaging is one ofthe methods used to diagnose LSS. Diagnosis by interpreting MR imagesrequires serious expertise, and it has been frequently studied by academics inrecent years because it is a system that assists the doctor with an objectiveapproach. This field of study is machine learning, which we can call the subbranchof Artificial Intelligence. Deep learning-based machine learning is verysuccessful in processing biomedical images such as MR. In this study, a modelthat performs 3-dimensional automatic segmentation on T2 sequence LumbarMR Images is proposed for the diagnosis of LSS. This 3D LSS segmentationstudy, according to our knowledge, has the feature of being the first in its fieldand will be an important resource for those who work in this field. In addition,with the proposed model, parts that cannot be fully opened in LSS surgicaloperations, especially in the nerve roots, can be fully determined beforehandwhich will ensure that the patient's complaints are completely eliminated afterthe operation. In MR images, a total of 6 classes were created and segmentationwas carried out, including the spinal disc, canal, thecal sac, posterior element,and other regions and background in the image, which are importantfor LSS. To measure the success of segmentation, the Intersection over Union(IoU) metric was calculated for each class. 3D segmentation success for the validationset in the dataset; Background (IoU = 0.83), Canal (IoU = 0.61), Disc(IoU = 0.91), Other (IoU = 0.97), Posterior element (IoU = 0.82), and ThecalSac (IoU = 0.81). The 3D automatic segmentation success rates obtained arequite high and show that a Computer Aided Diagnosis system can be createdin LSS diagnosis.