Content-based medical image retrieval using a novel hybrid scattering coefficients - bag of visual words - DWT relevance fusion

Image content analysis plays a major role in image classification, retrieval, and indexing together with object and scene recognition. Numerous image content descriptors are proposed in the literature, but their high computational costs and lower-performance scores make them inappropriate for content-based medical image retrieval (CBMIR) for large medical image datasets. To overcome these drawbacks, a novel hybrid Scattering Coefficients - Bag of Visual Words - Discrete Wavelet Transform (SC-BoVW- DWT) relevance fusion algorithm is proposed for effective CBMIR. For preprocessing, resizing and contrast limited adaptive histogram equalization (CLAHE) are carried out. Then scattering transform (ST), BoVW and DWT are applied to extract texture features, visual features, and low-level features of the preprocessed images respectively. A hybrid Grey Wolf Optimization - Particle Swarm Optimization (GWO-PSO) approach is used for optimal feature selection. Finally, image fusion (IF) is carried out with a relevance fusion based Euclidean distance technique. Experiments based on three standard medical computer tomography image databases namely, EXACT-09, TCIA-CT and NEMA-CT, are carried out. The proposed hybrid method outperforms the existing techniques in terms of precision, F-score, and recall values. Improvement in the average rate of precision (ARP), average rate of recall (ARR) and F-score values of 6.02%, 2.82% and 3.57% respectively, for the EXACT-09 dataset and 3.55%, 1.84%, and 2.12% respectively for the TCIA-CT dataset are observed compared with the existing Scattering Transform- canonical correlation analysis vertical projection (ST-CCA(v)). For NEMA-CT an improvement of 0.21% (ARP) is obtained compared with the existing Histogram of compressed scattering coefficients (HCSC).