Content based Biomedical Image Retrieval Models using Multi Resolution Texture and Region based features

Abstract

Content based image retrieval (CBIR) has a crucial role in medical science. It leads to earlier and automated detection of various diseases, retrieval of diseased images from databases and labeling of images in medical databases. In the thesis, an extensive survey to identify different types of cancers in human body has been performed and various CBIRs for such cancerous images are explored and their pros and cons have been discussed. In this context, a number of CBIRs have also been suggested. One such technique that has been suggested in this thesis is based on the extraction of wavelet transform coefficients up to five levels followed by possible dimensionality reduction using Probabilistic Principal Component analysis. For the extracted features, two classifiers a feed forward Artificial Neural Network (ANN) and K Nearest Neighbor (KNN) have been used for classification purpose. The performance is evaluated using two publicly available databases and compared with existing parallel methods in literature. The maximum increase in accuracy has been observed for ANN classifier up to 28.42% over another wavelet based technique for classification. A texture based CBIR using Local Mesh Peak Valley Edge Pattern (LMePVEP) has been enhanced by incorporating directional features in five directions 0o , 45o , 90o , 135o and 180o . The modified LMePVEP feature vector has been used for image retrieval purpose and its performance is evaluated for three datasets including one real dataset of Deen Dayal Upadhyay Hospital, Delhi. Three existing techniques are used to compare its performance in terms of Average Retrieval Rate (ARR), Average Retrieval Precision (ARP) and Accuracy. An increase of 5.26% in terms of accuracy has been observed with respect to the conventional LMePVEP