A Novel Content-Based Similarity and Softmax Method for Medical Image Retrieval

摘要

The content-based image retrieval (CBIR) has been greatly facilitated with the explosion of research in computer science. Efficiency and Accuracy of retrieval result become essential to medical diagnosis and treatment. In this research, a novel content-based similarity and softmax (CBSS) method is proposed for medical image retrieval. The multi-class classifier based on softmax function is introduced to improve the performance of medical image classification. The output of the classifier, with a raw image after transforming into high-dimensional feature space as input, provides an unbiased estimation in probability of raw image labels, which ensures the prediction accuracy and computational efficiency. In addition, the traditional similarity metric is also enhanced in our work to evaluate the medical images. The valuable visual signatures that extracted from the structure of the image sample are applied to calculating the distance as an approximation of similarity. The optimization strategy is capable of producing statistically significant improvements by combining the classifier and enhanced similarity matching. Experiments on a collection of medical images demonstrate the expected retrieval precision and recall rate of our approach comparing with Local Binary Pattern (LBP), Uncertain Location Graph Retrieval (ULGR) and Local Diagonal Extrema Pattern (LDEP). The expectation of retrieval precision of CBSS is about 54.5%, 16.2% and 10.14% higher respectively, which demonstrate the effectiveness of our proposal.