Intervertebral Disc Segmentation Using Machine Learning

Abstract

Disc degenerative changes are the most common cause of lower back pain. Treatment for this acute or chronic pain is physiotherapy or spine surgical procedures. These procedures include laminectomy or discectomy in which affected discs are surgically treated (removed). And for this purpose, disc location, size and shape are the main prerequisites. Initially surgeons used to rely on manually segmented data by Radiologists. Machine learning has revolutionized the medical field with the ability of making computers learn the common trends about disease patterns and predict pathological diagnosis in a very robust way. Many methods of machine learning have been developed for localization and segmentation of different anatomical structures, tumors or other pathologies and also used for histological studies of human body tissues. Moreover, several advancements in machine learning has increased the accuracy of diagnosis like in deep learning, densely connected convolutional neural networks have proved to be the accurate way of segmentation with high dice score The main pupose of this study is to segment the intervertebral discs automatically using densely connected network integrated into U-Net model in order to make computer aided diagnosis and surgical planning using MRI images of different modalities. The data set used is IVDM3Seg and taken from MICCAI 2018 challenge provided at grandchallenge.org. Segmentation models are trained on different MRI modalities given in the dataset and finally all the trained models are ensembled to get a satisfactory output. Final model resulted in 99.1% dice accuracy score which came out to be far better than previously used techniques.