This project aims to apply novel Bayesian Rule Learning (BRL) algorithms to estimate their feasibility for predictive modeling using cardiac MRI-derived markers. Structural and functional image markers will be extracted from publicly available cardiac- MRI datasets using standard and in-house developed image processing and extraction algorithms. The measurements from these markers will be input to the BRL algorithms to learn discriminative markers for the diagnosis and treatment of ischemic cardiomyopathy.
Period: April 2013 - April 2018
We intend to use the data for developing automatic algorithms to segment the left and right ventricles (blood pool and myocardium). We plan to learn image features that may help us in segmenting different regions and also exploit mutual context information between the left and right ventricles in the form of relative distance and angles. We will use random forest (RF) classifiers (trained on manual segmentations) to generate probability maps and use graph cuts for obtaining segmentation labels. In another approach we want to exploit the contextual relationship between the left and right ventricle by learning their relative orientation angle relationship. The relationship will be integrated into a cost function and optimized using graph cuts to get segmentation labels of different pixels.
Period: October 2012 - October 2013
We propose a framework for developing an automated CAD to detect wall motion abnormalities based on a novel classification system that is a hybrid between the statistical shape model (SSM) and Bayesian networks. The goal of this hybrid classification system is to overcome the weaknesses of each individual classification method and retain its strengths, thus increase the classification accuracy. Such a CAD system is expected to be a reliable assistance in reducing the inter- and intra-observer variability, and improving the diagnostic accuracy, especially for those less experienced cardiologists. Therefore, the first step to achieve our objective is to collect data and learn from the data, e.g. the relations of the regional and global clinical features (such as ejection fraction) with different pathology and healthy individual. Moreover, if the ground truth at segment level (The score that retrieved from visual wall motion scoring technique) is provided, then we can also use the clinical features of the data to train the Bayesian networks method, and finally to test on our hybrid system based on the trained SSM and Bayesian networks.
Period: November 2012 - November 2014
In a step towards aiding the design of left ventricle (LV) acting medical devices the Trinity Centre for Bioengineering has developed a new method of image analysis for interpreting the LV chamber and its walls. This new method converts volumetric measurements into detailed two dimensional data which is highly desired when characterising the working space for a medical device in the LV. The method works with the aid of an auxiliary plane view that moves sequentially through 3D computer models that are generated from Dicoms showing the short axis. We are working with the CAP database as it provides a large number of cases from which we can calculate mean values based on sex, age, disease state, etc.
Period: July 2012 - August 2012
Our group has recently developed a new MRI tagging sequence that produces radial, rather than cartesian (grid or line) tags. Due to the nature of the radial tag generation process the patient table in the scanner must be moved to an optimal location relative to the isocenter of the magnet in order to obtain the best radial tagging pattern. This table-shift distance is calculated from the centroid of patient's left ventricular blood pool in a short-axis image and the image orientation (normal). We propose to use the CAP database to prospectively determine the percentage of patients for whom an optimal table shift can be accommodated without exceeding the table-shift maximum and gradient linearity. To do so we will manually extract the LV centroid (if the data isn't available in CAP) and automatically extract image orientation information from the DICOM header. These parameters from the CAP database will be used to prospectively estimate the optimum table position for a large patient population.
Period: September 2011 - September 2013
The fast Radial Basis Function (RBF) method is particularly aimed at registering different medical image datasets typically used during surgical guidance, where both accuracy and speed of registration are of importance. The method displays subsecond performance and has been successfully tested on the registration of standard-sized data with small to medium discrepancies. The purpose of this validation study is to assess the method on registering datasets which display larger discrepancies hence displaying larger deformations during registration.
Period: 18 April 2011 - 16 December 2011
This project develops new mathematical techniques for direct shape mapping, registration and statistical analyses and provides tools for association of subject phenotypes with shape measures derived as cardiac biomedical imaging markers.
Period: April 2011 - 30 June 2011
This project aims to develop and apply new machine learning techniques in cardiac imaging based on shape, motion and appearance markers for efficient image indexation and retrieval from large 4D MR datasets.
Period: 01 April 2011 - 31 December 2014