Improved Convolutional Neural Networks in Content Based Image Retrieval System for Medical Image Analysis

Abstract

Content based image retrieval (CBIR) system is one of the significant research arena in digital image processing and computer vision. Image retrieval is a challenging task and CBIR is used to reduce the challenges by ascertaining the semantic gap between human queries and images in datasets. In particular, medical image datasets are rich in information content, CBIR works similar to human visual saliency mechanism and the visual features are considered for automatic searching process in a simple manner. Without using features like text or metadata, CBIR supports to manage the dataset with searching and organizing capabilities. This research work addresses the issues in content based medical image retrieval by investigating and proposing an efficient method to support CBIR systems using improved convolutional neural network. The semantic gap between the low level features and high level features are reduced using hash code generation process for medical image data and the improved convolution neural network is used to process the selected features to perform retrieval operation. To improve the retrieval accuracy of CNN, sparse representation is used in the proposed work.  Parameters such as classification performance, retrieval efficiency are experimentally verified and compared with conventional CNN and artificial neural network models. The proposed CBIR model has an average efficiency of 98.7% compared to other models.