Logo and Text Removal for Medical Image Retrieval

Abstract

The amount of visual medical information being produced in large hospitals is exploding. Most University hospitals produce Millions of images per year (Geneva Radiology: 20'000 images per day). Currently, the access to these images is most often limited to an access by patient identification. Sometimes search by text in the radiology report or in the DICOM headers is possible. Still, all the implicit information potentially available through the image and the accompanying case report is discarded in this case. Content—based visual data access is based on direct visual properties of the images that are extracted automatically from all images in a database. This delivers objective features for searching images but the features are commonly on a very low semantic level (colour histograms, simple texture analysis such as wavelet filter responses). Another problem that especially occurs in medical teaching files but also in routine images is text and logos around the main object in the image. For retrieval this is mainly noise that can have a negative influence on retrieval quality. In our approach, we extract the main object from the image by removing logos that are added to the images as well as frames around the images and text fields or other elements that are not needed. This is mainly based on properties of the text that occurs on the images, and especially of the logo of the university hospitals of Geneva. Frames around the images are removed reliably. First results show that the retrieval quality can be augmented well with such an approach. Especially queries with relevance feedback deliver much better results as the query is more focused. Proper, quantitative evaluation on a large data set is still missing but will be performed shortly.