Natural Language Processing Versus Content-Based Image Analysis for Medical Document Retrieval

Abstract

One of the most significant recent advances in health information systems has been the shift from paper to electronic documents. While research on automatic text and image processing has taken separate paths, there is a growing need for joint efforts, particularly for electronic health records and biomedical literature databases. This work aims at comparing text-based versus image-based access to multimodal medical documents using state-of-the-art methods of processing text and image components. A collection of 180 medical documents containing an image accompanied by a short text describing it was divided into training and test sets. Content-based image analysis and natural language processing techniques are applied individually and combined for multimodal document analysis. The evaluation consists of an indexing task and a retrieval task based on the "gold standard" codes manually assigned to corpus documents. The performance of text-based and image-based access, as well as combined document features, is compared. Image analysis proves more adequate for both the indexing and retrieval of the images. In the indexing task, multimodal analysis outperforms both independent image and text analysis. This experiment shows that text describing images can be usefully analyzed in the framework of a hybrid text/image retrieval system.

Figure 1

Sample document in the test corpus. An English translation of the text is provided by the authors for illustration purposes: Caption: “François, 1 1/2 month: Abdomen without preparation.” Paragraph: “Based on the child’s age and the fact that he experienced projectile vomiting, a diagnosis of pyloric stenosis can be made. An emergency abdomen without preparation is ordered for François; no evidence of gastric dilatation is visible.”

Figure 2

Processing multimodal biomedical documents for information retrieval.

Figure 3

Set of images in the test corpus matching the query “1121-115-700-400”.

Figure 4

Text and image analysis performed to assign IRMA codes to a test document. Text analysis (solid line and box) involved applying a dictionary and using the training collection to retrieve the 5 nearest neighbors (5-NN) for the text portion of a query document. The image analysis (dashed lines and boxes) used either the training collection or the IRMA database to retrieve the k nearest neighbors (k-NN) based on content-based image retrieval (CBIR) features.

Figure 5

Sample IRMA-MeSH equivalences made into dictionary entries (using MeSH hierarchy).

Figure 6

Sample IRMA-MeSH equivalences made into dictionary entries (adapting to IRMA specificity).