minimal-redundancy-maximal-relevance (mrmr)

On the one hand, the development of modern medicine has made it possible to store medical images, and on the other hand, due to the daily increase in the storage of such data, it has also made their management and recovery difficult. Considering that medical images are used as a powerful tool in the early diagnosis of most diseases, providing a powerful system that can retrieve images with similar content from the growing volume of medical images is very effective in control and treatment. In this article, a medical image retrieval system based on Siamese neural network consisting of two convolutional sub-networks with 13 layers is presented. To reach the optimal subset of deep features extracted by Siamese, the Minimum Redundancy-Maximum Relevant (mRMR) technique has been used, and after binary hashing of the features, similar images are retrieved using Hamming distance. Although the proposed model is capable of retrieving a variety of gray scale medical images, two types of lung images have been used to evaluate it, including CT scan images of Covid-19 patients in the CT-COV database and X-ray images of pneumonia patients in the Pneumonia database. The results indicate that the proposed method in the Covid database has been able to achieve an average precision of 93.83% and 92.73% in 5 and 10 retrieved images respectively, and an average precision of 100% in the pneumonia database, which is compared to previous methods have been able to improve the retrieval of lung images.