This paper describes the participation of the SNUMedinfo team at the two retrieval tasks (Ad-hoc image-based retrieval and Case-based retrieval) in the ImageCLEF 2013 medical task. For the ad-hoc image-based retrieval task, we submitted 1 baseline textual run using query likelihood model in Indri search engine, and 4 visual runs utilizing various image features implemented in Lire image retrieval library. For the case-based retrieval task, we submitted 1 baseline textual run using query likelihood model in Indri search engine, and 9 textual runs utilizing external corpus (MEDLINE) for expansion term inference. Our method ranked first in the case-based retrieval task.
We submitted 1 baseline run (SNUMedinfo11) using unigram language model with Dirichlet prior smoothing[2, 3]. Only figure caption field in document were indexed. Experimental results are described in Table 1.
We submitted 4 visual runs using various features implemented in Lire image
retrieval library[4], respectively. These features are : 1) the auto color correlogram
feature which, as the faster and more efficient version of the color correlogram[
There are several query images given per each query. We draw ranked list per each query image, and then combined them using Borda-fuse methods[9].
Experimental results are described in Table 2.
We submitted 1 baseline run (SNUMedinfo10) using unigram language model with Dirichlet prior smoothing. Title, abstract and fulltext field in document are indexed. The queries are stopped at the query time using the standard 418 INQUERY stopword list, case-folded, and stemmed using Porter stemmer.
For other 9 textual runs (SNUMedinfo1~9), we utilized external corpus (MEDLINE) for robust and effective expansion term inference. We leased 2013 MEDLINE/PubMed Journal Citations from the U.S. National Library of Medicine, composed of roughly 22 million MEDLINE citations. Our method can be summarized in the following steps (k and w is parameter).
(1) External corpus (MEDLINE) documents are indexed. Title, abstract and Mesh descriptors fields are indexed. (2) Using original query, MEDLINE documents were retrieved.
a. Among retrieved MEDLINE documents, k documents whose publication type is “Case Reports”, are selected from the top rank. b. In selected k documents, MeSH descriptors whose “MajorTopicYN” attribute value is “Y” are collected for expansion term. (3) Original query is augmented with expansion terms. Expansion terms are weighted by parameter w. (4) Expanded query is applied on target corpus.
Indri query example is described below. #weight ( (1-w) #combine( original query terms )
w #combine( expansion query terms ) ) #weight ( 0.7 #combine(48 old woman right cheek swelling blocked nasal passage … ) 0.3 #combine(maxillary neoplasms myxoma odontogenic tumors … ) ) We tried diverse parameter combinations (k : 1, 5, 10; w : 0.15, 0.3, 0.45) to evaluate the effect of parameter value on the performance. Experimental results are described in Table 3.
Overall, our methods brought 20~30% performance gain (in MAP) compared to the baseline method, and ranked first in the case-based retrieval task.
For the ad-hoc image-based retrieval task, we submitted 4 visual runs using several basic image features without applying additional techniques. These runs showed unsatisfactory performance (MAP range from 0.0019 to 0.0043), which means more advanced technique is required for this task.
For the case-based retrieval task, we submitted 9 textual runs (SNUMedinfo1~9) utilizing external corpus (MEDLINE) for expansion term inference. This method showed effective performance (20~30% gain in MAP), and is also robust across diverse parameter values.
This work was supported by the Seoul National University Brain Fusion Program Research Grant.
6. References 1. 2. 3. 4. 5. 6.
Aslam, J.A. and M. Montague, Models for metasearch, in Proceedings of the 24th annual international ACM SIGIR conference on Research and development in information retrieval2001, ACM: New Orleans, Louisiana, USA. p. 276-284.