|
|
|
|
|
|
|
|
| About Javed Mostafa (University of North Carolina) |
|
Dr. Mostafa is the Frances Carroll McColl Term Associate Professor at the University of North Carolina at Chapel Hill, USA,
with a joint appointment in information science and in the Biomedical Research & Imaging Center (a medical school entity). He is the Co-Director of the Biomedical Informatics Core at the Translational Clinical Sciences Institute (CTSA-funded Institute) and he is the Director of the Laboratory of Applied Informatics Research -- both based in UNC. His main area of research is information retrieval, with a particular focus on developing effective computational functions for analysis, visualization, and
personalization of biomedical information. He is also involved in developing educational programs in health informatics and digital libraries. He is an associate editor of the ACM Transactions on Information Systems.
|
|
Medical Image Content-based Retrieval for Clinical Decision Support
Javed Mostafa (University of North Carolina at Chapel Hill)
Significant technical advances have been made in image processing and
analysis. However, their applications to medical image retrieval,
particularly in operational contexts, have been limited. Our goal is to
conduct classification and indexing of medical images in near real-time,
i.e., soon after the capture process is completed. A related goal is to
develop high-accuracy retrieval and analysis functions over a dynamic image
collection which is continuously expanding. There is a difficult trade-off
involved in accomplishing these goals, mainly requiring a balance between
the representation process and access-latency. Generally, a representation
process improves as more time is dedicated to it; however, it leads to
increased access-latency.
We created a research platform to analyze new and different approaches for
reaching a balance in the trade-off. The platform, called ViewFinder, was
designed to support image manipulation by specialists in the domain of
Alzheimer's disease. Currently, the system supports several retrieval functions based
on textual metadata. Under development is a content-based image retrieval
function to allow a physician to use any patient's image in the
collection to retrieve images of other patients who are approximately in
the same disease stage. As a training collection we are using the
pre-classified image data set from the Alzheimer's disease Neuroimaging
Initiative (ADNI) at UCLA. We designed ViewFinder based on a multi-level
architecture, permitting convenient integration of modules for feature
generation, classification, and retrieval.
We have been experimenting with multiple algorithms for feature generation.
Two algorithms, namely Discrete Cosine Transform and 2-D Splines, have
produced the best results. We have also implemented a classifier based on
Support Vector Machines (SVM) and used a Euclidean distance measure to
arrive at ranking for the retrieved images. In this presentation, I will
discuss the case of a trade-off between representation and access in the
context of functions supported in ViewFinder and present experimental
results to demonstrate how a balance between representation and access
potentially can be achieved.
|
|
|
|
|
|
|
|
|
|
|
