A Content-Based Medical Image Retrieval Method Using Relative Difference-Based Similarity Measure

Content-based medical image retrieval (CBMIR) is a technique for retrieving medical images based on automatically derived image features. There are many applications of CBMIR, such as teaching, research, diagnosis and electronic patient records. Several methods are applied to enhance the retrieval performance of CBMIR systems. Developing new and effective similarity measure and features fusion methods are two of the most powerful and effective strategies for improving these systems. This study proposes the relative difference-based similarity measure (RDBSM) for CBMIR. The new measure was first used in the similarity calculation stage for the CBMIR using an unweighted fusion method of traditional color and texture features. Furthermore, the study also proposes a weighted fusion method for medical image features extracted using pre-trained convolutional neural networks (CNNs) models. Our proposed RDBSM has outperformed the standard well-known similarity and distance measures using two popular medical image datasets, Kvasir and PH2, in terms of recall and precision retrieval measures. The effectiveness and quality of our proposed similarity measure are also proved using a significant test and statistical confidence bound.