Detecting and Identifying the Lumbar Spine Deformities using Deep Neural Networks

Abstract

Lumbar Spine plays a very important role in our load transfer and mobility. Vertebrae localization and segmentation are useful in detecting spinal deformities and fractures. Understanding of automated medical image is of main importance that will help the doctors in handling the time consuming manual or semi manual diagnosis. Our thesis presents the methods, that will help the clinicians to grade the severity of the disease with confidence, as the current manual diagnoses by different doctors has dissimilarity and variations in the analysis of diseases. In this research we are discussing the lumbar spine localization and segmentation which help for the analysis of lumbar spine deformities. Lumber Spine is localized using YOLOv5 which is the fifth variant of YOLO family. It is the fastest and the lightest weight object detector. Mean Average Precision (mAP) of 0.975 is achieved by YOLOv5. To diagnose the Lumbar Lordosis, we have correlated the angles with region area that is computed from the YOLOv5 centroids and got 74.5% accuracy. Cropped images from YOLOv5 bounding boxes are passed through Hed U-Net which is combination of segmentation and edge detection frameworks, to get the segmented vertebrae and its edges. Lumbar Lordortic Angles (LLA) and Lumbosacral Angles (LSA) are found after detecting the corners of vertebrae using Harris Corner Detector with a very less mean error of 0.29˚ and 0.38˚ respectively. Lumbar spine is segmented using three deep neural models, the best results are achieved by U-Net with highest dice coefficient score (DC) and Intersection over Union (IOU). This thesis compares the results of three deep learning models that are Fully Convolutional framework, U-Net and SegNet architecture.