Font Size: a A A

Computational models of stereo vision with applications to optimal stereo x-ray imaging

Posted on:1999-07-23Degree:Ph.DType:Dissertation
University:Georgetown University Medical CenterCandidate:Frye, Richard EugeneFull Text:PDF
GTID:1468390014967738Subject:Biology
Abstract/Summary:
Although early cancer detection facilitates successful treatment, cancer remains a leading killer. Both lung and breast cancer can be radiographically identified. Modern digital image processing provides radiographic feature enhancement and automatic diagnostic software. Neural networks detect spiculated breast lesion with better accuracy than well trained radiologists. However, such image processing does not provide any novel information.;A radiologist can visualize a 3D virtual image of imaged tissue if stereo X-ray image-pair are acquired. Someday, algorithms will reconstruct the 3D structure of the imaged tissue. A 3D reconstructed image allows regions of interest which lack obscuring superimposed overlying and underlying structures to be produced and examined in detail. Stereo image-pair can be optimally acquired and displayed by adjusting the geometric parameters of the capture and display subsystems. Methods for calculating the values of these geometric parameters have not been outlined.;The geometry of stereo image-pair acquisition and display subsystems require optimization in order to account for factors such as object size, depth resolution and human psychophysiology. Subsystem parameters can be manipulated to produce stereo image-pair which capture the total object depth while providing the best depth resolution. When an acquisition and display system are combined, the maximum depth resolution is dependent on the display subsystem parameters, whereas the relative depth resolution is dependent on the acquisition parameters. Capturing crossed-only disparity depth space produces the best depth resolution. Examples of utilizing the derived equations for medical imaging are given.;I chose to simulate the neurophysiology of the mammalian visual cortex by building a model of stereopsis mechanisms using Gaussian derivative basis functions. A function set was distributed across a range of multiple spatial frequencies by extending the derivative order or scaling the first two or four Gaussian derivatives. Both position and disparity information could be derived from simulated stimuli by (1) adding and subtracting the output of one function set or (2) combining information from two spatially offset function sets. The scaled function set utilizing the first two derivatives combined with addition and subtraction demonstrated a stereo acuity pattern consistent with the human visual system.
Keywords/Search Tags:Stereo, Depth resolution
Related items