Using Media Fusion and Domain Dimensions to Improve Precision in Medical Image Retrieval

Analysis of biomedical and health queries: Lessons learned fromTRECandCLEFevaluation benchmarks

“…In addition, the analysis reveals that half the queries belonging to task T3 have more than 15 words whereas queries belonging to the other tasks contain a maximum of 16 words. This is explained by the length of the queries (despite the fact they were just a few) provided in CLEF case base retrieval track, created from the teaching file Casimage (Radhouani et al., ); they allow simulating the situation where a clinician diagnoses a difficult case described through a long description including patient demographics, limited symptoms, test results, and image studies. The same overall trend is observed for the number of query concepts (noted by LgC ( Q )), ranging from 0 to 12 for all the queries across the different tasks, but does not exceed 7 for tasks T1 and T2, and 4 for task T4.…”

“…We particularly highlight that the difficulty of the TREC Genomics 2003 queries can be explained by the relevance pooling set up the year the IR task started (first year in 2003): documents were assessed as relevant if a Gene Reference into Function (GeneRIF) was available for the document. Giving the fact that the number of available GeneRIFs was small, the number of relevant documents is obviously underestimated (Radhouani et al., ). We can notice moreover that for the CLEF case base IR tasks, the proportion of easy queries since 2012 is quite low (23% for 2012 and 29% for 2013), with a threshold fixed to 0, and a large number of queries having a null MAP score (77% and 71% of null values, respectively).…”

“…It has been organized by the U.S. National Institute of Standards and Technology (NIST, http://www.nist.gov/) (Voorhees, ) since 1992 and provides to participants different controlled tasks' specifications with related textual document collections, queries, relevance assessments, and other metadata depending on the task being evaluated. Historically, in the area of medical‐ and health‐related IR, TREC has evolved since 2003 through the following tasks: TREC Genomics: the TREC Genomics task was one of the largest and longest running challenge evaluations in biomedicine; it was launched in 2003 to answer first the challenge of managing and retrieving medical literature in order to identify potential interactions between genes, diseases, and other biological entities (Radhouani, Kalpathy‐Cramer, Bedrick, Bakke, & Hersh, ). This task models the setting where a genomics researcher entering a new area presents a query to a search engine.…”

“…The system should return the documents that are relevant for the considered query. Although the document collections were subsets of MEDLINE medical database abstracts in 2003 and 2005, the document collection included in 2006 full‐text HTML documents from 49 journals that were electronically published via Highwire Press (Radhouani et al., ). TREC Filtering (Robertson & Hull, ): this track seeks to measure the ability of an IR system to select relevant documents that fit persistent users' needs represented by profiles. It is worth mentioning that we specifically used in our study the medical data set provided within this track for an ad hoc retrieval purpose, not a filtering one.…”

“…Domain‐based indexing relies on the use of semantic resources such as UMLS, MeSH, ICD‐10, SNOMED, etc. In our study, we make use of the following forms of text (either document or query) indexing: The MeSH terminology : our choice for this terminological resource is based on the following: (a) previous work clearly shows that it is the most used general resource in the biomedical domain (Radhouani et al., ; Stokes et al., ), so it allows us to make the results issued from our study comparable to other ones issued from prior studies of the literature review; (b) because MeSH entails general biomedical terms, its coverage is wider compared to subdomain specialized terminologies such as Gene Ontology. Considering the use of different medical document data sets relying on different subdomain knowledge (gene, medical, etc.…”