Logo and text removal for medical image retrieval

1
Logo and text removal for medical image retrieval

• Workshop BildVerarbeitung in der Medizin,
  15.3.2005

Henning Müller University of Geneva Service of
Medical Informatics
2
Overview

• Medical image retrieval
• Our medical teaching file casimage
• Concrete problems for retrieval
• Logos and text in the images
• Removal steps (using itk)
• Evaluation of removal results
• Better retrieval?
• Conclusions

3
Content-Based Medical Image Retrieval

• Search for images in databases by visual means
  (textures, colours, forms) instead of text
• Navigate in very large databases in a different
  way
• Complementary to text (no replacement)
• Main applications
• Teaching find similar images to an example
• Also with a different diagnosis, but visually
  similar
• Research optimize the case selection
• Diagnostic aid in specialized domains
• lung CT, pathology,
• Specific applications DICOM header correction

4
An image retrieval example
5
medGIFT and casimage

• http//www.sim.hcuge.ch/medgift/
• Medical image retrieval system based on the GNU
  Image Finding Tool (GIFT)
• http//www.casimage.com/
• Teaching file internally and on the web with
  60000 images internally 9000 on the web
• Multi-lingual, mediocre annotations
• Connected to PACS and contains applications to
  anonymise images, etc.
• All sorts of images are stored in there

6
Examples of problems
University logo added to many images
Text in the images
Specific problems
Large regions with no information
7
Goals

• Fully automatic extraction of the main object in
  the image
• Usable on an extremely large variety of medical
  images that are present in our teaching file
• Small rate of image parts removed by error
• No default for background colour
• Better navigation and retrieval in the teaching
  file by visual means

8
Characteristics of the logo and text

• Fine structures of unconnected components
• Logo is always at the same position and same size
• Logo is white (350-400 pixels)
• Can be detected and removed with an erosion
• Grey square on lower right corner
• Detected through thresholding
• Median filter already removes much of text, plus
  removal of small unconnected components

9
Using itk to remove unwanted structures

• Itk makes development and tests quick
• Includes routines to open images, etc.
• Removal of specific structures (logo, grey
  square)
• Prior specific knowledge
• Smoothing (median filter)
• To remove small structures
• Edge detection
• To detect the main structure
• Thresholding
• Removal of small unconnected elements

10
Removal steps
11
Problems

• Images with slow changes
• Images with very fine structures (text)

12
Results of the removal

• 500 images segmented from the collection were
  analysed manually for the segmentation quality
• 185 image is fine, no work necessary
• 204 main object is extracted as wanted
• 105 not everything was extracted as wanted
• 4 too much was removed

13
More results
14
Retrieval performance an example
15
Retrieval performance more systematic

• Subjective performance gets better
• Use of ImageCLEF competition data
• Slightly worse results with segmentation
• MAP 0.3757 vs. MAP 0.3562
• Results with relevance feedback do get better
• Reasons
• Ground truth of ImageCLEF is not complete
• Features for shape detection work worse on
  segmented images as there is no clear border
  anymore
• Feedback is needed to focus the query

16
Conclusions

• Fully automatic object extraction is possible
  with a low error rate
• Specific problems need to be analyzed manually
• Hand-made solutions
• For indexation, more work is needed
• Create a small area around the object to have a
  border for retrieval
• Extract only object not the entire bounding box
• Logo and text removal treats several problems of
  automatic indexing of medical image databases