Our projects for medical image retrieval
Lung Image Retrieval: the Talisman project
A couple of steps have to be taken for proper lung image retrieval.
The first steps can not be done in an automatic way. A medical
doctor (MD) has to supply the necessary information. Only like
this, a proper evaluation of the retrieval quality is possible.
Tools have to be developed to make this interactive process as
easy as possible for the MD. At the University Hospitals of the
Geneva a set of tools is the CasImage
program that is a collection of medical cases including medical
images, especially for teaching but that could also be used for
case-based diagnostics.
- Choose slices of the HRCT that represent the disease well;
- link a diagnosis with a number of slices;
- give a verbal description of the abnormalities of the tissue;
- mark the region(s) in the slices that correspond best to the
pathology.
Evaluation of retrieval systems and medical reference databases
Currently we have a clear lack of image databases available free
of charge for evaluation purposes that include ground truth (or
a gold standard). To really evaluate the performance of retrieval
systems such databases and gold standards need to be created at
several scales, for varied (PACS-like) databases where the goal
needs to be to find visually similar images and specialized databases
such as those containing lung high-resolution lung CTs, where
the goal can be a diagnostic aid. Some of the available databases
on the internet include the database of the Casimage
project containing more than 8200 images free of copyright. This
databases is in the accessible in the MIRC
standard that offers some more free radiological resources. The
European Federation for Medical Informatics has also founded an
iniative to create a
Reference Image Database. The National Library of medicine
(NLM) offers access to some images such as those of the
visible human project.
Combination of visual and textual features
Image retrieval by visual features can only be a complement to
textual search for images and not a replacement. Best results
can be expected when combining visual and textual cues. This has
to our knowledge never been done in an in an integrated form with
medical images. A web demonstration with a combination of visual
and textual features for museum images can be found at Monash
University.
|