This paper describes the participation of the IPL team in the ImageCLEF-2010 campaign. Our research group has participated in the ad hoc task the case-based and the modality tasks of image retrieval based only on textual data. To reach our aim we performed a quantitative evaluation of Lucene search engine using several state of the art similarity functions. Our runs performed moderately well using the MAP and the early precision (P10) metric.
This paper presents the participation of the IPL at the CLEF 2010 Image retrieval task. The main goal of the medical ImageCLEF task is to improve the retrieval of medical images from heterogeneous and multilingual document collections containing images and text. Queries are formulated with sample images and some textual description explaining the search goal.
Image retrieval systems do not currently perform as well as their text
counterparts [
This year's image collection contains approximately 77,000 images. Some records had only one image while others had many images. This situation makes a purely text-based retrieval of images problematic however, since although only one image from a record containing many images might be relevant to a query based on the content, all the images of the record will be equally retrieved, reducing in this way the precision of the search.
Most of our e ort this year was concentrated in a quantitative evaluation of a search engine under several state of the art similarity functions.
Our retrieval system is based on Lucene [
To achieve this goal we incorporated within the lucene search engine three similarity functions, as will see in more details in the next section, namely the BM25, the axiomatic similarity function and the document pivoted normalization function.
Each gure in the database represented in XML format as the following simple example.
simple record < record > < pmid > 12345 < =pmid > < articleurl > http : ==www:someurl:com=somearticle:html < =articleurl > < caption > somecaption < =caption > < f igureid > f ig:xyz < =f igureid > < f igureids > < f igureidss > f ig:xyz:a < =f igureidss > < f igureidss > f ig:xyz:b < =f igureidss > < f igureidss > f ig:xyz:c < =f igureidss > < =f igureids > < meshterms > < meshterm > meshterm1 < =meshterm > < meshterm > meshterm2 < =meshterm > < meshterm > meshterm3 < =meshterm > < meshterm > meshterm4 < =meshterm > < =meshterms > < =record >
We note that several gures referred to the same image with the same title and same captions, so we decided to unify those records into one, with one header " gureid" and the rest gures as relevant ones. Thus we have two di erent elds for the gures of the case ( gureid, gureids), and the rst one plays the role of the header gure and the rest the role of relevant gures. Moreover the title and caption of the uni ed gure were merged into a single eld named "caption" that contains all the textual information that accompanies each gure.
For the indexing with the Lucene, we used our own Analyzer which performs tokenization, removes stop words, short words (less than 2 characters), long words (more than 50 characters), transforms each word to lower case, and performs stemming using the Porter stemmer. 2
The ad hoc task involves retrieving relevant images using the text associated to
each image query. For this task we investigated four similarity functions with
the lucene search engine: the default similarity function[
X tf (t; Q) t2Q^D
N df (t) 0:35
tf (t; D) tf (t; D) + 0:5 + 0a:5vgjDdlj where tf (t; Q) is the number of occurrences of term t in query Q, df (t) is the number of documents that contain term t, jDj is length of document D, and avgdl is the average document length in the collection. 3. Document Pivoting Normalization
This approach uses the default scoring function of lucene but tf (t; D) which is given by the square root of the term frequency is now estimated by the relation (4) tf (t; D) = 1 + log(f req(t; D)) 1 + log(avgF req(D)) lwehnegrteh NavogrFmre=q(pD()1 =slojpUe1D) jpPi1vott2+dslfopreeqjU(Dt;jDa)n,dUpDiv=ott=he jDd1ijsPcredt2eDtjeUrmDjs in D, The results out of these four functions from the CLEF-2009 image-collection are presented in table 1.
In table 2 are presented the Ad Hoc runs we submitted and a short description for each one.
In the AdHoc version of the Medical Retrieval, the unit of retrieval is a gure. A gure is identi ed by the gureid. The goal of the retrieval is to correctly retrieve the most relevant gures for each one of the 16 queries- gures. To complete this task we create indexes with the Lucene search engine using only the textual information of the gures (title + caption).
Moreover as we can see in the runs with su x TCM the documents are expanded with the MeSH terms. In this case we have the two searchable elds. In that case we weight the two elds by ratio 0.9/0.1 in favor of "caption" over the "mesh". The scoring function is given by: score(Q; D) = 0:9 score(Q; DC ) + 0:1 score(Q; DMeSH ) (5) where the eld DC contains the title and caption and the DMeSH the MeSH terms. In the Case-based version of the Medical Retrieval, the unit of the retrieval is not an image but a case. Cases are identi ed by a PubMed id ("pmid" tag in the xml format). Unfortunately not all cases are marked with a "pmid", so the identi er is the "ArticleURL" tag of the xml format. In other words, a case is pointed by a link in the web of an article that describes it.
A simple fusion method has been implemented to obtain a simple list of relevant gures. The total score of each relevant document is calculated by the sum of the scores for all the retrieved gures in the document: score(Q; D) = X score(Q; F ) (6) Finally, the documents are sorted by their nal fusion score.
Table 5 summarizes the runs we submitted and a short description for the Case-based task.
Similarly in this case we use two types of indexes. For the rst type, only the eld "contents" (title+caption) is searchable. At the retrieval phase we get the top 1000 results and iterate over them to add the score of the results with the same "articleurl". This procedure produces the run named "ipl aueb casebased CT".
In the second type of indexes, there is an additional searchable eld "mesh", we performed 5 runs. The number in the run's name denotes the weight given in the "mesh" eld. A value 0.1 means a weight 0.1 for the score of the "mesh" eld and 0.9 for the score of the "contents" eld.
Table 6 contains the top 10 performing participations in the case-based track based on the MAP metric. IPL was at the 28th position at the sequence with MAP equal to 0:1228. For the task of modality classi cation, a list of 2,390 records was given for training. The goal was to correctly classify the 2,620 records of the test set into one of the eight modalities: "CT" , "GX" , "MR" , "NM" , "PET" , "PX" , "US", "XR".
We used two classi ers on this task: a modi ed Perceptron-type algorithm
we have described in previous work [
Our baseline textual system performed quite well, with a MAP of 32%. We will continue to improve our image retrieval system by adding more image tags using automatic visual feature extraction and heuristics. There was only a minimal improvement in performance with the use of the image modality. As future research we plan to improve the pseudo-relevance feedback strategy. 6