Similarity Searching of Medical Image Data in Distributed Systems: Facilitating Telemedicine Applications

Introduction

Healthcare centers and hospitals collect daily a huge amount of medical data, including images that capture structural (e.g., Computed Tomography (CT), Magnetic Reasoning Imaging (MRI)) and physiological/functional information such as functional MR imaging (fMRI) etc., as well as clinical assessment data such as lab tests and physician evaluations. In the information age we live in, an important issue is to be able to extract knowledge from these data through analysis in order to obtain a deeper understanding of normal and disease states. Moreover, advances in networking and communication technologies can support the efficient exchange of medical data between organizations. Thus, combining the advantages brought by these developments can assist diagnosis in a more effective way.

The above mentioned advancements in both communication and information technologies in medicine had as a result the birth of the emerging field of telemedicine. In such applications, medical information is transferred through the Internet or other networks, for the purpose of consulting or medical knowledge sharing, making less necessary the transfer of patients. Many researchers are interested in the integration of telemedicine applications. Besides technological issues, these applications should take into consideration the nature of information to be transmitted between the sites, the quantity of information to be transferred (Gemmill, 2005; Harnett, 2006) and security and privacy as well (Appari, 2009; Ruotsalainen, 2010).

In this study, we propose a method for performing similarity searches of medical image data in a distributed environment. These searches are performed over a distributed index that we build. The distributed index is composed of indices of individual images. Those indices point to the original images. Each image’s index contains a feature vector from which only a low resolution image can be obtained, the IP address of the source node that holds the original image and a unique name of the image. The distribution of the image’s index is based on a multi-resolution analysis of the images and on a set of reference images known by all nodes. For the multi-resolution analysis, we use the wavelet decomposition that is very popular in medical imaging and is used, except for feature extraction, in denoising, image compression, etc. Although we do not distribute the images that are stored in the local databases of the sites that participate in the network, but rather a pointer to them that is based on features that are extracted from the original image, our system has the ability to access other clinical data and physicians’ evaluations on similar clinical cases, assisting in effective medical decision making. This can be achieved by issuing a query image. The query will be routed to the node that stores similar image’s indices to the query image, then if one wants more information about an image can submit a request to the node that holds the original image. Our framework provides a way of communication between the clinicians and the sites that hold the original images (by providing information about the healthcare center (i.e., its IP address) and the original image (i.e., a unique name that the center has assigned to the image). As a result, the physicians can access the clinical data if the healthcare center’s protocols about security and confidentiality issues, allows it. Considering performance and scalability issues we also pay attention to load balancing. Although the images are stored in the healthcare centers that participate in the network this does not mean that all nodes of the network will hold approximately the same number of images’ indices after building the distributed index (especially the index that is built using the content of the distributed data). This problem is known as load balancing and has as a result some peers of the network to answer more queries than others. Here, we deal with this problem introducing a scheme to assign to the nodes similar numbers of images’ indices.

To test our method, we use images of the same modality such as fMRI images. Taking into consideration the fact that a medical record of a patient may contain data from multiple modalities, our framework can easily be extended because of the fact that is based on image’s extracted features. In our implementation, we use peer to peer networks because of their wide use in data sharing. The proposed framework has the advantage of keeping bandwidth requirements to a minimum by reducing the amount of data that need to be transferred, while medical decision making is not affected by the data reduction.