Segmentation and Classification of Disc Hernia in Magnetic Resonance Images using Deep Learning

Abstract

Localization of lumbar discs in magnetic resonance imaging (MRI) is a difficult task,mainly due to the disc variance in size,shape,number,and appearance. This paper proposes a deep learning methodology for automatic segmentation and classification of disc herniation. The dataset used in this research included publicly available database Lumbar Spine MRI Dataset obtained from Mendeley Data,combined with images obtained from patients from the Clinical Centre of Kragujevac,Serbia. Total number of images was 1169 images in sagittal view and 557 images in axial view. The methodology includes several steps starting from segmentation of disc,bounding box cropping and enhancement of disc region,after which the classification based on convolutional neural network (CNNs) is performed (healthy,bulge,central,right or left herniation for axial view and healthy,L4/L5,L5/S1 level of herniation in sagittal view). Results show 0.87 accuracy for classification in axial view images and 0.91 accuracy for sagittal view images. The obtained results represent the advancement in comparison to the state-of-the-art results,where mostly binary classification (healthy or herniated disc) is investigated. Future research will focus on increasing the dataset size,investigation of other deep neural network architectures,as well as employing transfer learning in disc hernia classification.