A computer vision system for the automatic assessment and identification of stenosis from angiograms

Date of Award




Degree Name

Doctor of Philosophy (Ph.D.)


Interdepartmental Studies

First Committee Member

Tzay T. Young, Committee Chair


The computer vision system developed in this study is aimed at the complete automation of the angiographic image analysis. The functional components of this vision system are: Low-level input analysis, segmentation, image representation, and intrinsic image computations and interpretation. The low level stage employs subtraction, conditional smoothing and intensity transformation. Subtraction is done by utilizing the whole sequence of images and by calculating minimum and maximum gray levels over time. Conditional smoothing and intensity transformation functions are used to clear out the noise and redundancy. The segmentation is a multiresolution and bidirectional method which consists of several stages of rule based algorithms. The first stage of the segmentation is a quad-tree based split and merge algorithm to which four rules are developed. Then, the resulting intrinsic image is passed through rule based merging and rule based melting processes. The rule based merging uses a region adjacency graph as a basis for decisions whereas the rule based melting algorithms utilize a minimum cost spanning tree.The angiogram is finally represented in terms of a binary image and the medial axis of the vessels constructed from the binary image. Based on these images dimensions of the vessels and the flow rate for every point on the medial axis of the vessels are computed. In addition, some roughness measures are calculated as well. Then, using results from the quantification of the dimensions the reliability of the system is determined. The roughness measures are used independently to localize the stenoses in the vessels. Combination of these stages generated a broader understanding of the internal structure of the vessels. Another rule based algorithm was developed to trace and identify the individual branches of the arterial tree in order to bring an understanding of how vessels overlay each other.The complete system was evaluated by three dynamic phantoms which were developed for this purpose. The segmentation stage was also evaluated by two clinical images. The results of the quantification of the dimensions of the vessels were found to be within acceptable error rates. The localization and identification of the stenoses generated very reliable results. The tracing algorithm was successful in identifying all the branches for the three dimensional phantom except for one section where three arteries were overlapping each other.


Health Sciences, Pathology; Computer Science

Link to Full Text