Application of computational intelligence in medical image analysis for spine related diseases

Abstract

Spine related diseases, such as intervertebral disc degeneration, lumbar spinal stenosis and cervical myelopathy, are very common all over the world. Medical images have been widely used in interpretation, diagnosis and prognosis of these spine related diseases. Currently, medical images are mostly analyzed with conventional analysis methods, such as subjective observation, manual measurements and conventional statistical tests, in studies of spine related diseases. These conventional methods are usually limited by subjective biases, time-consuming process and poor ability in handing large number of features. Sometimes, these limitations cannot be neglected, greatly hindering the application of medical image in clinical practice. Computational intelligence (CI) would somehow be an alternative. Proper CI method can integrate numerous features for a comprehensive analysis and achieve interpretation, diagnosis and prognosis of spine related diseases in an objective and time-saving way. Therefore, this project would like to investigate the way to apply CI methods in medical image analysis for spine related diseases. Five specific applications covering image interpretation, diagnosis and prognosis of spine related diseases were studied, including: (1) Three-dimensional measurement of cartilaginous endplate (CEP) based on ultrashort time-to-echo magnetic resonance imaging (MRI) technique. Conventionally, the status of CEP could only be evaluated in two dimensional images; the proposed three-dimensional measurement could provide more information for evaluation of status of CEP. (2) Classification of symptomatic and asymptomatic lumbar spinal stenosis. MRI features could only be individually analyzed with conventional methods; a new classification model was proposed to combine these MRI features by carefully designed feature selection, feature transformation and classification algorithms. The classification accuracy had been greatly improved with CI method. (3) Segmentation of lumbar spinal tissues from axial T1-weighted MRI image. Features extracted from axial T1-weighted MRI images had been proved to be valuable in diagnosis of lumbar spine disorders. Automatic segmentation of lumbar spinal tissues could help extract accurate and reliable features. A hierarchical segmentation framework was proposed in this project. This segmentation framework utilized a rough-fine strategy and worked in multiscale image spaces. Segmentation results that were comparable to manual extraction ones were acquired with this framework. (4) Prognosis of cervical myelopathy (CM) based on diffusion tensor imaging (DTI) technique. DTI was supposed to be an effective tool for prognosis of CM. However, no significant process was made using conventional analysis methods. This project proposed a new type of DTI features and used machine learning algorithm to analyze the features. The prognostic accuracy had been greatly improved with the proposed prognostic model. (5) Evaluation of performance of DTI metric maps in region of interest (ROI) drawing for DTI-based studies on cervical spinal cord. With CI methods, the performance was quantitatively evaluated by examining their abilities in separating the gray and white matters. CI methods have been successfully applied in these applications, where conventional methods don’t work well. This finding suggests the feasibility and effectiveness of CI methods in medical image analysis for spine related diseases.