Last updated on February 13th, 2014
So I have officially finished my Master’s work 🙂 and I’ve written up a more extended abstract of the thesis with less technical terms below. This hopefully can give you some concise intuition on the work without the technical details. This page is meant to be a summary of all things related to this work and links to other relevant articles that I wrote.
The thesis is available here and is entitled:
Spinal cord segmentation and disability prediction in multiple sclerosis using novel optimization and machine learning methods
This thesis examines a) how to segment the spinal cord from 3D magnetic resonance images (MRI) using computer-aided methods, and b) how to predict the level of clinical disability in multiple sclerosis (MS) patients using spinal cord segmentations and corresponding MRIs.
Spinal cord segmentation is important since spinal cord atrophy has been shown to correlate with the progression of MS. Accurate cord measurements may be used as biomarkers to monitor the progression of diseases and the effectiveness of therapies. Current methods to segment the cord are laborious and not robust to cord and MRI changes.
Predicting the disability of the patient using the spinal cord segmentations and MRI information is important as this may assist clinicians in monitoring the disease. As well, the features used to model disability may give insights into how MS affects the spinal cord.
To segment the spinal cord we propose two different approaches (you can see a recorded video on these approaches in my Talk on spinal cord segmentation). Our first approach represents the spinal cord by a list of medial axis coordinates and shape parameters that describe the cross-section of the spinal cord. We model the cross-sectional shape using a statistical shape model (principal component analysis) built from manually segmented spinal cords. For a particular cross-sectional shape, we examine how well the created shape fits with the image information by considering the gradient information and intensity information in the MRI. We optimize for the medial path and shape parameters using an A* minimal path search to finds the optimal segmentation that fits with the information in the MRI. For more details, please check out this post and poster on Globally optimal spinal cord segmentation using a minimal Path in High Dimensions.
Our second approach to segment the spinal cord takes a machine learning approach to detect and regularize the spinal cord. We train a series of decision forests to classify between the spinal cord and the background on a per-voxel basis. The predicted values of the previous classification is passed as input to the next classifier as per the auto-context algorithm. At each stage of auto-context, we consider the connected components of the predicted voxels to represent candidate shapes. We then extract “global features” from this candidate shape (e.g. volume) and use these features as input into the classifier. This improved the spinal cord segmentations when compared with our first approach. The poster for this work can be viewed in my post on Augmenting auto-context with global geometric features for spinal cord segmentation.
To predict the clinical disability in patients with MS, we extract features (e.g. intensity, ellipse fitting) from patients’ spinal cord segmentations and corresponding MRIs. We use these features to build linear and non-linear models that are trained to predict the level of physical disability in MS patients. Using leave-one-out cross-validation, we test how well our created models can predict the disability of patients using only the information found in the MRIs and spinal cord segmentations. We show that non-linear models with two features (minimum distance to the boundary of the cord from the centroid and the major axis of fitting an ellipse) provide superior indication of disability level when compared to linear models using only the spinal cord volume as suggested by the literature. The slides that I did for this talk can be viewed in this post on Novel morphological and appearance features for predicting physical disability from MR images in multiple sclerosis patients.