Medical image retrieval systems have gained high interest in the scientific community due to the advances in medical imaging technologies. The semantic gap is one of the biggest challenges in retrieval from large medical databases. This paper presents a retrieval system that aims at addressing this challenge by learning the main concept of every image in the medical database. The proposed system contains two modules: a classification/annotation and a retrieval module. The first module aims at classifying and subsequently annotating all medical images automatically. SIFT (Scale Invariant Feature Transform) and LBP (Local Binary Patterns) are two descriptors used in this process. Image-based and patch-based features are used as approaches to build a bag of words (BoW) using these descriptors. The impact on the classification performance is also evaluated. The results show that the classification accuracy obtained incorporating image-based integration techniques is higher than the accuracy obtained by other techniques. The retrieval module enables the search based on text, visual and multimodal queries. The text-based query supports retrieval of medical images based on categories, as it is carried out via the category that the images were annotated with, within the classification module. The multimodal query applies a late fusion technique on the retrieval results obtained from text-based and image-based queries. This fusion is used to enhance the retrieval performance by incorporating the advantages of both text-based and content-based image retrieval.