Image Area Reduction for Efficient Medical Image Retrieval
MetadataShow full item record
Content-based image retrieval (CBIR) has been one of the most active areas in medical image analysis in the last two decades because of the steadily increase in the number of digital images used. Efficient diagnosis and treatment planning can be supported by developing retrieval systems to provide high-quality healthcare. Extensive research has attempted to improve the image retrieval efficiency. The critical factors when searching in large databases are time and storage requirements. In general, although many methods have been suggested to increase accuracy, fast retrieval has been rather sporadically investigated. In this thesis, two different approaches are proposed to reduce both time and space requirements for medical image retrieval. The IRMA data set is used to validate the proposed methods. Both methods utilized Local Binary Pattern (LBP) histogram features which are extracted from 14,410 X-ray images of IRMA dataset. The first method is image folding that operates based on salient regions in an image. Saliency is determined by a context-aware saliency algorithm which includes folding the image. After the folding process, the reduced image area is used to extract multi-block and multi-scale LBP features and to classify these features by multi-class Support vector machine (SVM). The other method consists of classification and distance-based feature similarity. Images are firstly classified into general classes by utilizing LBP features. Subsequently, the retrieval is performed within the class to locate the most similar images. Between the retrieval and classification processes, LBP features are eliminated by employing the error histogram of a shallow (n/p/n) autoencoder to quantify the retrieval relevance of image blocks. If the region is relevant, the autoencoder gives large error for its decoding. Hence, via examining the autoencoder error of image blocks, irrelevant regions can be detected and eliminated. In order to calculate similarity within general classes, the distance between the LBP features of relevant regions is calculated. The results show that the retrieval time can be reduced, and the storage requirements can be lowered without significant decrease in accuracy.