Spinal Imaging Data Analysis & Clinical Parameter Extraction Using Transformers

Abstract

Human spine is a complex structure that plays a vital role in the movement, protection, and support of the body so it is very important to follow proper spine bio-mechanics to avoid any unwanted effect on body. Spinal diseases can cause compression or pulling the nerve roots, which can lead to radicular symptoms like back pain or leg pain. On the other hand, it may cause deformities which are most common at C4-C7 and L4-S1 level. Localization of vertebra bones that make up the spine is key in spinal disease diagnosis such as calculating cobb angles, shape detection, detecting vertebra fractures and other abnormalities. In this paper, we have covered four modules, first, for the vertebrae localization we used detection transformer to localize 68 corner points, Secondly, we have used a SegFormer to do the segmentation of the spine. Thirdly, center profile of the spine was generated using center point technique for localization and morphological thinning for segmentation, In the final step of shape analysis process, we take the profile of spine to calculates the features and classify the data into normal, Single-bend (C-shaped) and Double bend (S-shaped) spine. DETR gives mAP value of 0.96 at 0.5 IOU threshold and SegFormer achieves a dice score of 0.93 in segmenting spinal images. For the classification of the data, we have used different classifier (SVM, RF, KNN and NB). We have obtained identical features from both Segformer and DETR techniques and compared the results. Features acquired from localization technique(DETR) yield better accuracy when using a random forest classifier.Random forest performs best for AASCE MICCAI 2019 dataset with an accuracy of 98.9%.