Medical Image Retrieval Based On Semi-supervised Learning

With the development of medical imaging technology, the number of medical images pro-duced by hospitals and research institutes grows enormously. How to effectively manage and analysis medical image database is a key issue in the biomedical domain. Content-based med-ical image retrieval searching clinical images with similar pathology according to their visual contents becomes an important tool for computed-aided diagnosis and medical research. How-ever, there are three main problems in existing medical image retrieval systems:firstly, it is hard to recognize and describe the medical ontology from image due to its complexity; second-ly, there is a semantic gap between low level features and high level semantics, so that using visual characteristics as a basis of similarity matching and measurement can not effectively represent user’s query semantic; finally, the process of retrieval ignores the medical annotation and its semantic concept in the medical case.To solve the above problems, this thesis discusses four critical aspects in medical image retrieval based on the graph-based semi-supervised learning theory:image retrieval combining visual semantic; relevance feedback model; multi-modality medical case retrieval; efficient-ly retrieval algorithm. The main contributions and innovative achievements of this thesis are following:(1) A medical image retrieval algorithm combining visual semantic and local features is proposed. This algorithm firstly embeds manifold assumption in the graph-based semi-supervised learning framework and represents a label propagation algorithm with density constraint, in order to obtain the membership degree of the query image as the visual se-mantic; secondly, the dense SIFT features is extracted from the image blocks to generate the visual words as the local feature; finally, a similarity measurement based on visual semantic and local feature is designed. Experimental results demonstrate that the pro-posed algorithm represents the visual semantic of images effectively, and achieve a better retrieval performance than single low level feature.(2) According to the feedback by the query from doctors, in this thesis we construct a rel-evance feedback model based on pairwise constraint propagation. The basic idea is to obtain pairwise constraints from user feedback and propagate them to the entire image set to reconstruct the similarity matrix, and then rank medical images on this new man-ifold. The proposed method solve the problem of small sample size and asymmetrical training typically in traditional relevance feedback. On this basis, the thesis represents a medical image retrieval framework using long-term feedback strategy, which can utilize historical information to update the retrieval model. Therefore, the proposed framework outperforms previous relevance feedback models.(3) For similar case searching in computer-aided diagnosis, we proposed a multi-modality medical case retrieval approach based on multi-graph semi-supervised learning. This approach regards medical images and annotations as compensatory multi-modality in-formation and defines a multi-graph fusion based semi-supervised learning framework to combine visual content and textual information, which aims to fuse the different modality features on the manifold according to the graph regularization model and obtain the final retrieval results by manifold ranking on the fusion graph. The experiments on several medical case datasets show that the proposed approach can effectively utilize the image and textual information for medical case retrieval.(4) A multi-feature oriented fast manifold ranking algorithm is represented. Because compu-tational complexity is a bottleneck of graph methods applying in real data processing, this thesis uses anchor graph to implement original graph reconstruction, in order to simplify the calculation complexity on the premise of holding manifold structure. The proposed retrieval model based on sequential fusion, not only reduces the complexity of graph con-struction and ranking calculation, but also promotes the performance of manifold ranking by multi-feature representation. Hence, it can significantly improve the query speed at the same time keep the retrieval accuracy.