A Robust Scheme of Vertebrae Segmentation for Medical Diagnosis

Abstract

Automated vertebrae analysis from medical images plays an important role in computer aided diagnosis (CAD). It provides an initial and early identification of various vertebral abnormalities to doctors and radiologists. Vertebrae segmentation and classification are important but difficult tasks in medical imaging due to low contrasts in image, noise and high topological shape variations in radiological scans. It becomes even more challenging when dealing with various deformities and pathologies present in the vertebral scans like osteoporotic vertebral fractures. In this work, we want to address the challenging problem of vertebral image analysis for vertebra segmentation and classification. In the past, various traditional imagery techniques were employed to address these problems. Recently, deep learning techniques have been introduced in biomedical image processing for segmentation and characterization of several abnormalities. These techniques are becoming popular in solving various medical image analysis problems due to their robustness and accuracy. In this research, we present a solution of vertebrae segmentation and classification problem with the help of deep learning approach. We present a novel combination of traditional region based level-set with deep learning framework in order to extract the shape of vertebral bones accurately; which would be able to handle the deformities in the vertebral bones precisely and efficiently. After vertebrae segmentation, we further extend the work to abnormal vertebrae classification with the help of deep learning approach. This novel framework would be able to help the medical doctors and radiologists with better visualization of vertebral bones and providing the initial automated classification of vertebrae to be normal or abnormal. The proposed method of vertebrae segmentation was successfully tested on different datasets with various fields of views. The first dataset comprises of 173 CT scans of thoracolumbar (thoracic and lumbar) vertebrae in sagittal view, collected from a local hospital. The second dataset comprises 73 CT scans of cervical vertebrae in sagittal view, also collected from a local hospital. The third dataset comprises 20 CT scans of thoracolumbar (thoracic and lumbar) vertebrae in sagittal view collected from spine segmentation challenge CSI 2014. The forth dataset comprises 25 CT scans of lumbar vertebrae in sagittal view collected from spine segmentation challenge CSI 2016. Lastly, we have utilized the same locally collected set of 173 iii CT scans of thoracolumbar (thoracic and lumbar) vertebrae and extracted in axial view to perform the segmentation task. For classification purpose, we have utilized the locally collected set of 173 CT scans of thoracolumbar (thoracic and lumbar) vertebrae as these include osteoporotic vertebral fractures in it. The details of these datasets have been presented in respective sections. We have achieved promising results on our proposed techniques. The evaluation of the segmentation performance on the datasets with various machines and field of views helped us to ensure the robustness of our proposed method. On validation sets of these datasets, we have achieved an average dice score of around 95% for vertebrae segmentation; and accuracy of above 80% for the vertebrae classification. The detailed results have been presented in the results section. These results reveal that our proposed techniques are competitive over the other state of the arts in terms of accuracy, efficiency, flexibility and time.