Date of Award
Master of Science
Eliot H. Winer
The advent of medical imaging technology enabled physicians to study patient anatomy non-invasively and revolutionized the medical community. As medical images have become digitized and the resolution of these images has increased, software has been developed to allow physicians to explore their patients' image studies in an increasing number of ways by allowing viewing and exploration of reconstructed three-dimensional models. Although this has been a boon to radiologists, who specialize in interpreting medical images, few software packages exist that provide fast and intuitive interaction for other physicians. In addition, although the users of these applications can view their patient data at the time the scan was taken, the placement of the tissues during a surgical intervention is often different due to the position of the patient and methods used to provide a better view of the surgical field. None of the commonly available medical image packages allow users to predict the deformation of the patient's tissues under those surgical conditions.
This thesis analyzes the performance and accuracy of a less computationally intensive yet physically-based deformation algorithm- the extended ChainMail algorithm. The proposed method allows users to load DICOM images from medical image studies, interactively classify the tissues in those images according to their properties under deformation, deform the tissues in two dimensions, and visualize the result.
The method was evaluated using data provided by the Truth Cube experiment, where a phantom made of material with properties similar to liver under deformation was placed under varying amounts of uniaxial strain. CT scans were before and after the deformations. The deformation was performed on a single DICOM image from the study that had been manually classified as well as on data sets generated from that original image. These generated data sets were ideally segmented versions of the phantom images that had been scaled to varying fidelities in order to evaluate the effect of image size on the algorithm's accuracy and execution time. Two variations of the extended ChainMail algorithm parameters were also implemented for each of the generated data sets in order to examine the effect of the parameters.
The resultant deformations were compared with the actual deformations as determined by the Truth Cube experimenters. For both variations of the algorithm parameters, the predicted deformations at 5% uniaxial strain had an RMS error of a similar order of magnitude to the errors in a finite element analysis performed by the truth cube experimenters for the deformations at 18.25% strain. The average error was able to be reduced by approximately between 10-20% for the lower fidelity data sets through the use of one of the parameter schemes, although the benefit decreased as the image size increased. When the algorithm was evaluated under 18.25% strain, the average errors were more than 8 y times that of the errors in the finite element analysis. Qualitative analysis of the deformed images indicated differing degrees of accuracy across the ideal image set, with the largest displacements estimated closer to the initial point of deformation. This is hypothesized to be a result of the order in which deformation was processed for points in the image.
The algorithm execution time was examined for the varying generated image fidelities. For a generated image that was approximately 18.5% of the size of the tissue in the original image, the execution time was less than 15 seconds. In comparison, the algorithm processing time for the full-scale image was over 3 y hours.
The analysis of the extended ChainMail algorithm for use in medical image deformation emphasizes the importance of the choice of algorithm parameters on the accuracy of the deformations and of data set size on the processing time.
Catherine Elise Peloquin
Peloquin, Catherine Elise, "Determination of critical factors for fast and accurate 2D medical image deformation" (2009). Graduate Theses and Dissertations. 11085.