Biomedical image query in Gaussian-modelled feature space employing GeoSOM with enhanced inverted indexing

Abstract

The paper investigates a novel visual concept-based biomedical image querying mechanism to manage with a wide range of medical images acquired under the similar class/modality. Initially, the approach converts RGB images to L*a*b* dimensions. L*a*b* triplets are then partitioned into sub-regions and subsequently modelled by an unskewed Gaussian function for image characterisation based on standardised moments, that represents visual features. The features are then blended to formulate visual concepts and later vector quantised by geodesic self-organising-map to build a codebook of concepts. A global relationship among the clustered concepts is further established through dual correlation matrices and used in a distance metric for matching the given query with target images. Additionally, an enhanced inverted indexing mechanism is introduced to reduce the search time of this query expansion framework. Systematic evaluation of retrieval results in a heterogeneous medical image collection of 30 000 images of different modalities, body parts and orientations, demonstrates 50% reduction in computation time and 20–30% increase in precision at each recall level in comparison with individual visual features. Further experimental analysis employing effectiveness, robustness and precision at 20 parameters validates the benefits of the proposed scheme.

Keywords:

• Gaussian characterisation
• Global correlation matrix (GCM)
• Global similarity matrix (GSM)
• Query expansion
• Vancouver distance
• Visual concepts