Our approach to the ImageCLEF medical case retrieval task consists of text-only retrieval combined with utilizing the Medical Subject Headings (MeSH) ontology. MeSH terms extracted from the query are used for query expansion or query term weighting. MeSH annotations of documents available from PubMed Central are added to the corpus. Retrieval results improve slightly upon full-text retrieval.
Medical case retrieval (MCR) is the problem of nding descriptions of diseases or
patients' health records (document corpus) that are relevant for a given
description of a patient's symptoms (query), as decided by medical experts. The
ImageCLEF medical task 2013 contains an instance of this problem aiming at fully
automatic retrieval, where the document corpus contains about 75,000
biomedical publications that should be queried for 36 symptom descriptions consisting of
text and diagnostic images [
The question whether and how text retrieval can be improved for medical
case retrieval is therefore an interesting research problem, where retrieval
performance is measured by mean average precision (MAP) or binary preference
(bpref) [
Our approach to the ImageCLEF MCR task is to expand a given query
by relevant MeSH terms in order to improve the average precision of fulltext
retrieval results. This idea is not new and has been applied with varying success
to the MCR problem [
The MeSH ontology consists of records representing the nodes of a tree structure.
A record describes a primary MeSH term and, among other information, a
number of synonyms (Figure 1). A parent node in the tree represents a more general
term than its child nodes. The child nodes of the root node (let us call them
toplevel nodes) are listed in Table 1. Following the approach of Diaz-Galiano et al.
[
To identify relevant MeSH terms of a query, we chose a simple approach: primary MeSH terms and synonyms were stored in a word-based inverted index after removal of hyphens and punctuation characters, and after lower-case ltering. After applying the same preprocessing to query words, they were looked up in the index and marked in the corresponding MeSH terms (primary or synonym). Finally, the ratio of marked to all words of a MeSH term was used to produce a ranked list of primary MeSH terms. By applying a threshold to ratio values, the nal list of relevant primary MeSH terms is obtained. For our experiments, we used a threshold of 0.8.
The original query was then expanded with the identi ed relevant primary MeSH terms and used for fulltext retrieval with one of two document indexes. The rst index was generated from the documents' fulltext including titles, abstracts, and gure captions. The second index was created from the same document text extended with annotated MeSH terms, which had been retrieved from PubMed Central3 in advance. Indexes were generated using Lucene4 3.6.2 with default token analyzer.
We also considered two variants of query expansion: (1) expansion by all
synonyms of a relevant primary MeSH term (as in [
The retrieval results of our methods are given in Table 2. Only 4 runs were submitted to ImageCLEF 2013 (indicated by the submitted run ID), the others were added after submission. All submitted runs did not improve upon the baseline run of the task organizers (run ID HES-SO-VS FULLTEXT LUCENE, 0.1791 MAP), and the submitted query expansion runs stayed below our own fulltext baseline. Query expansion with all synonyms of relevant primary MeSH terms (see variant (1) Section 2) reduced average precision dramatically. However, the query term weighting approach (see variant (2) in Section 2) increased MAP by 1.4% to 0.1838 with respect to our fulltext baseline. All our runs, though, stayed substantially below the best textual MCR run submitted to ImageCLEF 2013 (run ID SNUMedinfo9, 0.2429 MAP).
To address the ImageCLEF medical case retrieval (MCR) task, we followed a text-based approach. We identi ed relevant primary MeSH terms in the query text and performed query expansion against two fulltext indexes (excluding and including MeSH annotations). The method did not improve mean average precision of retrieval performance. However, a variant of query expansion amounting to query term weighting of primary MeSH terms already present in the query could slightly improve retrieval performance.
Strategies and parameters of the presented approach could be modi ed, e.g.
identi cation of relevant MeSH terms, using other top-level nodes and relations
in the MeSH ontology, and improving query term weights. Additionally, further
work will apply other known query expansion methods using external knowledge,
like pseudo-relevance feeback or MeSH term co-occurrence [