Comments: hadar
ContactPerson: huiluo@cse.buffalo.edu
### Begin Citation ### Do not delete this line ###
%R 2001-07
%U /u0/csestaff/stock/dissertation.prn
%A LUO, Hui
%T KNOWLEDGE-BASED IMAGE UNDERSTANDING AND CLASSIFICATION SYSTEM FOR MEDICAL IMAGE DATABASES
%D July 30, 2001
%I Department of Computer Science and Engineering, SUNY Buffalo
%K Medical images databases, Knowledge-based, image understanding,
image classification, feature extraction, snake model, object-oriented
representation
%X The understanding of medical images involves the identification of
anatomical structures in images, the establishment of the
relationships among the structures and the matching them with
pre-defined anatomical models.  The closer this match, the better the
image is understood. Hence, the development and use of anatomical
knowledge models are critical to the effectiveness of the system. In
this thesis, we will present a new knowledge model for medical image
content analysis based on the object-oriented paradigm. The new model
abstracts common properties from different types of medical images by
using three attributes: description, component, and semantic graph,
and specifies its actions to schedule the detection procedure,
properly deform the model shapes to match the corresponding anatomies
in images, select the best match candidates, and verify the
candidates¡¯ spatial relations with the semantic graph defined in the
model. The advantage of the new model includes: 1) It presents a
robust representation ability to represent the object variability, 2)
It provides a tighter bond between features and methods, and 3) It
supports the complex image analysis and works on both normal and
abnormal images.  Based on this model, a knowledge-based medical image
classification system is developed. The system assigns each input
image with one or more most similar models by performing a two-stage
processing. The first stage, coarse shape classification, focuses on
the global shape features of objects in the processed images. The
second stage performs a detail matching between extracted image
features and model specified properties. The output match results
imply a set of models which are most possible to be existed in the
processed image. To further confirm them, an evaluation strategy is
used to reject those unreasonable results and inference the high
confidence match models.  To improve the detection accuracy, we also
proposed a new region model which combines both region and edge
information into image segmentation, and reformulate the traditional
snake model and level set model with it. As expected, the reformulated
models demonstrate robust ability to deal with the ambiguity in images
and overcome the problem associated with the initialization.  The
performance of the system has been tested on different types of
digital radiographs, including chest, pelvis, ankle, elbow and
etc. The experimental results suggest that the system prototype is
applicable and robust, and the accuracy of the system is nearly 70% in
our image databases.