Detecting Disc Herniation in Segmented Lumbar Spine Magnetic Resonance Images using Distinct Features

Abstract

There is a need to develop a disc herniation decision support system,as lumbar disc herniation is one of the most prevalent causes of intervertebral disc diseases,accounting for 90% of all spine surgeries. Image processing techniques and artificial intelligence algorithms can be efficiently used to describe a strong relationship between the MRI of the lumbar spine disc and final diagnosis. This study presents a methodology for automatic segmentation of the region of interest (ROI) - disc area,after which distinct features are extracted and can be used for diagnosing disc herniation on axial MR images. U-net convolutional neural network is used for ROI segmentation,after which features such as moments,eccentricity,equivalent diameter,eigenvalues of the inertia tensor and ratio of bounding rectangle area and disc area are calculated. Additionally,centroid distance function was used to visually differentiate between the herniated and non-herniated discs. The methodology was based on the number of peaks on the graph plot of distance between the disc centroid and points on the edges of the disc. The results achieve dice coefficient of 0.961 and IOU of 0.925 for segmentation on axial test images,while eigenvalues of the inertia tensor proved to be very descriptive in differentiating herniated and non-herniated discs. The results from these tests can be regarded as an early step toward establishing a fully automated system for identifying lumbar spine disc herniation in future work.