medGIFT:

1
medGIFT Content-Based Access to Medical Image
Databases
Henning Müller Division of Medical
Informatics University Hospitals of
Geneva 26.6.-28.6.2003
2
Content-based access to data

• Large amounts of visual, digital data become
  available
• Internet
• PACS (12000 images per day at the radiology HUG,
  Geneva)
• Goal Access based only on visual features only
• Extraction completely automatic
• Grey levels (local, global, distribution)
• Textures (glocal, local)
• Forms of segments
• Show me images that are similar to
• Query by image example(s)
• No semantics in features, only complement to text

3
Example
Emphysema
Emphysema
?
Micro nodules
Macro nodules
4
Image retrieval in medicine

• Based on exact matches of data base fields
• Only rarely similarity or full text search is
  offered
• Commonly there are errors in DICOM headers
• Spelling errors in medical records
• Various ways of saying the same thing
• Visual access to images (current projects)
• To classify images into anatomical groups
• For long-term storage
• Diagnostic support
• Not a replacement of text

5
GIFT GNU Image Finding Tool

• http//www.gnu.org/software/gift/
• Demonstration at http//viper.unige.ch/
• Freely available (Open Source), communication
  interface (MRML, Multimedia Retrieval Markup
  Language)
• Easy integration into projects (open interfaces)
• Easy adaptations and specializations (modular
  structure)
• Visual features used
• Color histogram, local color features at
  different scales
• Gabor as texture description at global and local
  scales
• Weighting methods and relevance feedback as in
  text retrieval
• Tf/idf-based, information theoretic

6
medGIFT

• Adaptations have been done to GIFT for the
  retrieval of medical images
• More texture measures (scales, directions)
• More grey levels (32-64 perform best),
  corresponds to our perception
• Reference databases are in progress
• Performance comparisons
• Proof of performance can gain acceptance
• Integration into applications
• Casimage,
• Specialization of the system
• HRCTs of the lung
• Dermatology

7
Screenshot GIFT
Query image
Similarity score
Choice of relevance
Link to the complete case With full-size image
8
Future projects

• Combination of visual and textual features
• Quality control, one for the other
• Improve retrieval quality
• Connections for multi-lingual retrieval
• Integration into the PACS
• Extremely large amounts of data
• Other problems Choice of slices,
• Use as a dignostic tool
• Step-wise process
• Evaluation and user tests
• Specialization
• HRCTs of the lung
• Dermatology