Abstract:

Background and Aim: Among the types of documents to store medical information and medical image, because of the non-invasive method used to aid in diagnosis, is of great importance and according to statistics, its use in addition to diagnosis, research and education is rapidly growing. Such growth makes the retrieval of these images from systems with large volumes of medical images, an important task. Therefore, there is a need to have systematic review of retrieving images. The objective of this study was to conduct a systematic analysis of image retrieval methods, approaches and systems in the literatures between 2006- 2016.

Materials and Methods: The major scientific databases (Springer, Science Direct (Elsevier), IEEE, PubMed Central, SID and Magiran) were searched, using keyword "Medical Image Retrieval" in English databases and combination of "تصویر پزشکی AND بازیابی " for Persian databases between 2006 and 2016. All papers reviewed using standard critical assessment and enrolled based on their quality. Then, the data in selected studies were extracted and classified. Papers about review and evaluation or an innovation in the retrieval of medical images and medical image retrieval systems, got accepted in the study. Also, papers about relevance feedback, introducing solutions with content-based, text-based or both approaches to promote precision and recall, got accepted as well. Retrieved papers, which were about non-medical images or not about retrieval, were excluded from the study. Papers that repeated another study or were duplicated altogether, excluded from study. Also papers with any access to full-text were excluded.

Results: The result of search was a total of 47,585 papers. By excluding according to stated criteria, sixty-six selected articles were analyzed in terms of approaches, learning methods and datasets that were used for evaluation. Most papers related to medical image retrieval had content-based approach (71%) and a hybrid content-based and text-centric approach (15%), respectively. The rate of supervised learning method was 28% and unsupervised method 24% and there was almost equal tendency toward them. In most articles (39%), the researcher team gathered the evaluation dataset and then 18% of the articles used ImageCLEFmed, the standard test data set. Also, indexing techniques in the literature were analyzed.

Conclusion: The result of this study shows that due to the widespread applications of medical images, research in the development of effective methods capable of increasing the precision, recall and the speed of retrieval, is growing. Despite this, the lack of available standard benchmarks is a challenge for researchers and the statistics indicate the need to conduct adequate investigations in this case.