<title>Intelligent data layout mechanism for high-performance image retrieval</title>

Abstract

Trends in medical imaging indicate that the storage requirements for digital medical datasets require a more efficient, scalable storage architecture for large-scale RIS/PACS to support high-speed retrieval for multiple concurrent clients. As storage and networking technologies mature, the cost of applying such technologies in medical imaging has become more economically viable. We propose to take advantage of such economies of scale in technology to provide an effective network workstation storage solution for achieving (1) faster display and navigation response time, (2) higher server throughput and (3) better data storage management. Full-field direct digital mammography presents a challenging problem in the design of digital workstation systems for screening and diagnosis. Due to the spatial and contrast resolution required for mammography, the digital images are large (exceeding 5K X 6K X 14 bits approximately equals 60MB per image) and therefore difficult to display using commercially available technology. We are developing clinically useful methods of storing, displaying and manipulating large digital images in a medical media server using commercial technology. In this paper we propose an Intelligent Grid-based Data Layout Mechanism to optimize the total response time of a reading by minimizing the speed of image access (data I/O time) and the number of data access requests to the server (queueing effects) during the image navigation. A Navigation Threads Model is developed to characterize the performance of many navigation threads involved in the course of performing a reading session. In our grid-based data layout approach, a large 2D direct-digital mammogram image is divided spatially into many small 2D grids and is stored into an array of magnetic disks to provide parallel grid-based readout services to clients. Such a grid- based approach not only provides fine-granularity control, but also provides a means of collecting statistical information about the distribution of Region of Interests (ROI) for a given image in the storage systems. Hence, it provides statistical rules to guide image navigation and guidelines for reorganizing the data layout within the storage server (replication of ROI blocks) dynamically; hence, better load balancing can be achieved and the overall image navigation throughput for the system can be maximized. Given the same buffer capacity and data access mode, this technique can statistically guarantee the maximum image retrieval time, and can scale-up easily without significant performance degradation. Throughout this paper, we assume that a high- speed network is used in our client/server model and network latency (data communication cost) is minimal compared to data I/O cost. In addition, the cost of reporting diagnostic results associated with the total response time is assumed to be negligible.