Deep learning based vertebral body segmentation with extraction of spinal measurements and disorder disease classification

Abstract

Assessment of medical images and diagnostic decision making of lumbar associated diseases by clinicians is invariably subjective, time consuming and challenging task. Presently, clinicians make use of either manual or semi-automated computer-aided tools to make relevant measurements for adding vote of confidence to their grading and evaluation. Lacking reliability and offering substantive dissimilarity once performed by different clinicians, these methods complicate the evaluation process. In an effort to support the decision making process of clinicians, in this paper we present a lumbar assessment framework with autonomous extraction of spinal measurements. Furthermore, an effort is made to address the challenges faced by clinicians while assessing disorders including spondylolisthesis and assessment of lumbar lordosis (LL) by proposing novel disease classification methodologies. For spondylolisthesis classification, we achieved an accuracy of 89% by using angular deviation metric whereas, 93% accuracy for determining adequacy/inadequacy in LL assessment through computation of area within enclosed lumbar curve region. Our framework involves semantic segmentation of vertebral bodies (VBs) using ResNet-UNet where we achieved DSC of 0.97 and IoU of 0.86. Subsequently, we achieved a statistically significant correlation coefficient � and encouraging mean absolute error (MAE) with clinicians’ grading for measurements involving lumbar lordotic angle (LLA), lumbosacral angle (LSA), VB dimensions and lumbar height. In addition to this, we have publicly released the dataset with all the clinicians markings. Keywords: Lumbar spine dataset, Spinal measurements, Automated disease classification