This paper describes the SINAI team participation in the ImageCLEF campaign. In this paper we only explain the experiments accomplished in the medical task. We have experimented with query expansion and the text information of the collection. For expansion, we carry out experiments using MeSH ontology and UMLS separately. With respect to text collection, we have used three di erent collections, one with caption and title, other with caption, title and the text of the section where the image appears, and the third with the full article. Moreover, we have experimented with mixed search, textual and visual search, using the FIRE software for image retrieval. The use of FIRE and MeSH expansion with the minimal collection (only caption and title) obtains the best results in the track.
This is the fourth participation of the SINAI research group at the ImageCLEF 2008 campaign, speci cally in the medical task.
The goal of the medical task is to retrieve relevant images based on an image query. This year, a new collection have been used. It contains images from articles published in Radiology and Radiographics including the text of the captions and a link to the HTML of the full text articles. We have downloaded articles from the web and constructed a new textual collection including the text of the article section where the image appears. Besides, for the experiments, we have created two groups of expanded queries, a group expanded with MeSH ontology1 and other group expanded with UMLS2.
For mixed experiment, we have used the list of retrieved images by FIRE3 [
The following section describes the new textual collection. In Section 3, we explain the expansion of the queries. In the next section, we comment the experiments carried out. Finally, conclusions and further work are presented in Section 5. 2
This year, the old collection [
To create the di erent textual collections, rst we have obtained the textual information by following the next steps: 1. Extract a list of articles URLs from the information of the collection given by the organizers.
The numbers of articles are lower that the number of images that contains the collection, because an article contains several images.
3. Filter the downloaded articles to extract di erent sections in the text: title, authors, abstract, introduction, etc.
The we have created three di erent collections. In these collections each document contains information about each image from the original collection. The information is di erent for each collection. These collections and the section include the following sections:
CTS: Contains caption, title and text of the section where the image appear.
3
One of the purposes of these experiments is to compare the performance of query expansion using
two di erent ontologies: MeSH and UMLS. Experiments with the MeSH ontology have been
carried out in the past [
On the other hand, the UMLS metathesaurus is a repository of biomedical ontologies and associated software tools developed by the US National Library of Medicine(NLM)5. It is built from the electronic versions of many di erent thesauri, classi cations, code sets, and lists of controlled terms used in biomedical literature and health services research. These are referred to as the source vocabularies of the Metathesaurus in UMLS literature. One of the source vocabularies is MeSH ontology.
To expand the queries we have used MetaMap program [
bpoc: Body Part, Organ, or Organ Component
diap: Diagnostic Procedure dsyn: Disease or Syndrome neop: Neoplastic Process
MetaMap gives two types of mapped terms: Meta Candidates and Meta Mapping. The difference between both mapped terms is that the second are the Meta Candidate with best score. For our expansion we have used the Meta Candidate terms, because these terms provide similar terms with di erences in the words. For example, the phrase "`chest CT image"' obtains following candidates: 793 Image (Medical Imaging fMSH,MTH,NCIg) [Diagnostic Procedure] 734 Chest CT (Chest CT fMTH, SNOMEDCT, RCD, SNM, SNMI, ICD9CM, MTHICD9, MDRg) [Diagnostic Procedure] 604 Thoracic (Dissecting aneurysm of the thoracic aorta fMTH, CCPSS, ICPC2, ICD10ENG, ICD9CM, MDR, NCIg) [Disease or Syndrome] 577 Breast (Breast fHL7V2.5, LCH, MSH, MTH, NCI, PSY, RCD, SNM, SNOMEDCT, UWDA, CCPSS, LNC, AOD, CSP, SNMIg) [Body Part, Organ, or Organ Component] 577 Breast (Entire breast fMTH, SNOMEDCTg) [Body Part, Organ, or Organ Component] 560 Mammary (Mammary gland fMTH, RCD, SNM, SNOMEDCT, UWDA, MSH, NCI, AOD, CSP, PSY, SNMIg) [Body Part, Organ, or Organ Component]
The rst number is the score, the next one is the metathesaurus concept. The preferred name for a metathesaurus concept appears asided in parentheses. Between braces are the di erent source vocabularies where the concept appears, and between square brackets are the di erent semantics types of the concept.
Prior to the inclusion of Meta Candidates terms in the queries, the term words are added to a set where repeated words are deleted. All words in the set are included in the query. 4
Our main goal is to investigate the e ectiveness of di erent expansions and di erent sizes in textual collections. Moreover, we have experimented with the in uence of mixing visual information with our results.
We have used three textual collections (CT, CTS, CTA) and three sets of topics: original,
expanded with MeSH and expanded with UMLS. Besides, we have mixed our textual results with
the visual results given by the organizers in order to obtain new results. The visual results have
been obtained with the FIRE software. To mix textual and visual results, we have used the same
algorithm that applied in 2007 [
We have submitted only 10 runs, because of the limits imposed by the organizers. The results of these runs are shown in Table 1:
The o cial baseline results obtain 0,2768 of MAP (Main Average Precission). Table 2 shows the top ten MAP obtained in all experiments for all of participants.
Experiment sinai CT Base sinai CT Mesh sinai CT Umls sinai CT Mesh Fire20 sinai CTS Base sinai CTS Mesh sinai CTS Umls sinai CTA Base sinai CTA Mesh sinai CTA Umls
This new collection has been used in the ImageCLEFmed2008. Similarly to previous years, the use of textual information improved the results of baseline visual results. In this case, the use of FIRE and MeSH expansion with the minimal collection (only caption and title) obtains the best results in the track.
The use of UMLS expansion obtains worse results than the baseline. Although UMLS Metathesaurus includes MeSH ontology in the source vocabularies, MetaMap adds, in general, more terms in the queries. The MetaMap mapping is di erent than MeSH mapping, therefore the terms selected to expand are di erent.
Another conclusion is that it is better to have few textual information but more speci c. Including all the section where the image appears is not good approach. Sometimes, a section contains several images, therefore the same information references di erent images.
Our further work was to obtain a more precise textual information by nding the phrase where a reference to the image exists, that is, by nding HTML tags that reference locally to the gure (for example: A HREF=\#F1") or syntactic structures of type \in gure 1 ...". Moreover, we expanded the queries with UMLS Metathesaurus using other algorithm distinct to that used by the MetaMap tool. We investigated new methods to expand with UMLS similar to the expansion with MeSH, that is, less terms but better for information retrieval. 6
This project has been partially supported by a grant from the Spanish Government, project TIMOM (TIN2006-15265-C06-03), and the RFC/PP2006/Id 514 granted by the University of Jaen.