SINAI at ImageCLEF 2009 medical task

Abstract

This paper describes the SINAI team participation in the ImageCLEF 2009 medical task. We explain the experiments accomplished in the medical retrieval task (ImageCLEFmed). We have experimented with query and collection expansion. For expansion, we have carried out experiments using MeSH ontology. With respect to text collection, we have used di®erent collections, one with only caption, and the other with caption and title. Moreover, we have experimented only with textual search, using the LEMUR toolkit as Information Retrieval system.

Keywords

• H
• 3 [Information Storage and Retrieval]
• H
• 3
• 1 Content Analysis and Indexing
• H
• 3
• 3 Information Search and Retrieval
• H
• 3
• 4 Systems and Software

Introduction

The following section describes the collections used and the expansion of collection and queries using MeSH ontology. In Section 3, we explain the experiments and obtained results. Finally, conclusions are presented in Section 4. 2

The textual expanded collection

In 2008 a new collection was introduced in this task. We have created di®erent textual collections using information of this collection in the web [ 1 ]. This year, the task has been separated in two subtask, one based in image retrieval (adhoc) and the other based in medical case retrieval. For these reason, we have created three di®erent textual collections, two collections to image based retrieval and one to case based retrieval. The collections are: ² C: Contains caption of image to use in image based retrieval. ² CT: Contains caption of image and title of the article to use in image based retrieval. ² TA: Contains title and text of the full article to use in medical case based retrieval.

This year we have experimented with expansion of the textual collection. Our initial aim was to expand the textual collection in the same way that the query expansion in experiment of last year [ 1 ]. Nevertheless, the time request needed for expansion of minimal textual collection with UMLS and MetaMap software has been excessive. For this reason, we only have used the MeSH ontology to expand collections and topics.

From the collections used in image based retrieval, we have created other two expanding the text using MeSH ontology. These new collections have been named CM and CTM respectively. This year we have not used the collection with full article to image based retrieval for two reasons. The ¯rst one is that we want to experiment the expansion of collection with minimal textual information. The second reason is that the expansion of a big textual collection requires a large amount of time. 3

Experiment Description and Results

This year we have two types of topics sets: adhoc and case based topics. We have expanded this topics set and we have obtained the next topics sets: ² t: Original topics set to use in adhoc retrieval. ² tM: Topics set to adhoc retrieval expanded with MeSH ontology. ² cbt: Original topics set to use in case based retrieval. ² cbtM: Topics set to case based retrieval expanded with MeSH ontology.

The name of the experiment includes the name of the group, the name of the collection (described in section 2) and the name of the topic set. The data set of the collection has been indexed using Lemur3 IR system, by applying KL-divergence weighing function and using PseudoRelevance Feedback (PRF).

Table 1 shows the main average precision (MAP) of image based retrieval experiments. Di®erent from what we expected, these results show that query expansion does not improve the results. The expansion using only the collection with more textual information (CTM collection) obtains the best results.

Table 2 shows the results of medical case based experiments. As we can see in this table the query expansion does not improve the results.

Experiments sinaiCt sinaiCTt sinaiCMt sinaiCTMt sinaiCtM sinaiCTtM sinaiCMtM sinaiCTMtM

MAP 0.2626 0.2605

Conclusions

This year we have used topic and the expansion of the collection. The topic expansion has been carried out in the same way as previous year. The expansion of the collections improves the results when the topics are expanded, too. However, the obtained results are not successful. At the moment, we are investigating the reasons of these unexpected results.

Acknowledgements

This work has been supported by the Regional Government of Andaluc¶³a (Spain) under excellence project GeOasis (P08-41999), the Spanish Government under project Text-Mess TIMOM (TIN2006-15265-C06-03) and the local project RFC/PP2008/UJA-08-16-14.

References

[1] D ¶ ³az- Galiano , M.C. and M.A. Garc ¶³a-Cumbreras and M.T. Mart ¶³n- Valdivia and L.A. UreaL¶opez and A. Montejo-R¶aez. Query Expansion on Medical Image Retrieval: MeSH vs . UMLS. Lecture Notes in Computer Science . Springer. 2009 .

[2] D ¶ ³az- Galiano , M.C. , Garc ¶³a- Cumbreras , M.A. , Mart¶³n- Valdivia , M.T. , Montejo-Raez , A. , and Uren~a-Lo¶pez, L.A.: SINAI at ImageCLEF 2007 . In Proceedings of the Cross Language Evaluation Forum (CLEF 2007 ), 2007 .

[3] MuÄller , H. and Kalpathy-Cramer , J. and Eggel , I. and Bedrick , S. and Rahouani , S. and Bakke , B. and Kahn , C. and Hersh , W. Overview of the CLEF 2009 medical image retrieval track . CLEF working notes. Corfu, Greece . 2009 .

[4] D ¶ ³az- Galiano , M.C. , Garc ¶³a- Cumbreras , M.A. , Mart¶³n- Valdivia , M.T. , Montejo-Raez , A. , and Uren~a-Lo¶pez, L.A.: SINAI at ImageCLEF 2006 . In Proceedings of the Cross Language Evaluation Forum (CLEF 2006 ), 2006 .