Thomas Hangartner (Advisor), Ping He (Committee Member), David Short (Committee Co-chair)
Master of Science in Engineering (MSEgr)
To allow automatic assessment of computed tomography (CT) images of long bones, the identification of the location of regions of interest is important. The distance between the distal and proximal styloids may act as reference points for bone length; however, the locations of the distal and proximal growth plates represent further important reference features. The current methods to locate these feature positions are manual. In this project, we attempt to find these feature positions automatically.
A CT data set was split into two subsets, a development set, and a validation set. We first extracted basic information from the stack and re-sequenced the slices from proximal to distal in the pre-task. A calibration phantom and its compartments are automatically located. Calibration equations are used to adjust all intensity information to a common scale. For segmentation, three methods were used to detect the region of interest (tibia and radius): Search-Box, Search-Outline and Center Compare-Outline. These methods track the bone of interest from midshaft to both distal and proximal styloids with an outline and extract information such as slice-based region of interest (ROI) area, intensity, intensity standard deviation, trabecular bone density, center of bone etc. Using this information, with the assumption that the area relationship between inner growth plate and midshaft won't change much, the position of the end of the growth plate directed to the midshaft is detected. Further, the position of the styloids relative to the growth-plate position is estimated.
The contour success rates for the ROI for Search-Box, Search-Outline and Center Compare-Outline were 81.9%, 84.7% and 91.5% for legs and 63%, 83% and 89.3% for arms, respectively. The success rates for the growth-plate position estimation for Search-Box, Search-Outline and Center Compare-Outline were 20%, 77.5% and 85% for the leg development set and 10.5%, 64% and 74% for the leg validation set. The success rates for the arm development set was 11%, 67% and 53% and for the arm validation set 6%, 50% and 45%, respectively.
The Center Compare-Outline is the best method for its high success rate for contour and feature position estimation. It doesn't rely on the orientation of the bone as the other methods do. The whole stack gets analyzed satisfactorily while only one threshold based on the center slice is used. One strength of our method is its automaticity. It only requires a small amount of manual input to assign the distal-proximal direction and to give each bone a seed point; the rest is done automatically. Sometimes there are errors in the growth-plate search process, so visual assessment of the result with possible manual correction is still needed.
Department or Program
Department of Biomedical, Industrial & Human Factors Engineering
Year Degree Awarded
Copyright 2011, all rights reserved. This open access ETD is published by Wright State University and OhioLINK.