Issue No. 01 - January/February (1998 vol. 13)
DOI Bookmark: http://doi.ieeecomputersociety.org/10.1109/5254.653224
<p>Referencing prior images of a patient is critical to a radiologist's ability to interpret the images generated as part of a radiological examination. During an image reading, a radiologist typically exercises a series of judgments to determine what prior images are relevant. These judgments are complex, practice-dependent, and fluid in that they may change over a short time. Many health care institutions support image retrieval by prefetching images according to some primitive rules of thumb (such as the same modality and anatomical part as the current image). However, these practices are not sophisticated enough to meet many reading scenarios, so radiologists typically spend considerable time searching for additional prior images.</p> <p>An alternative is intelligent patient-image retrieval using either a knowledge-based system- in which a knowledge base of image-selection rules is manually engineered-or automated learning. For most organizations, the cost of engineering a knowledge base customized to the radiologist or institution may be prohibitive. The sidebar describes these limitations in more detail. We believe a more economic alternative is an AI-based mechanism that learns patient-image-retrieval heuristics. Such a mechanism would be key to the broad implementation and long-term maintenance of intelligent and adaptive image-retrieval systems.</p> <p>In this article, we describe the radiologist behaviors associated with patient-image retrieval and a backpropagation neural net-work that learns patient-image-retrieval heuristics. We compared the learning performance of this network with benchmarks from a system that uses an engineered knowledge base as well as with a real-world image-prefetching practice. We also investigated the network's ability to tolerate imperfect (noisy) data. Results show that, with fine-tuning, the image-prefetching accuracy of the network is comparable to that possible with an engineered knowledge base. The network also acceptably tolerates the noisy and incomplete input data characteristic of patient-image retrieval in this context.</p> <p>These results are significant because they demonstrate that AI-based learning techniques such as ours can provide intelligent information systems with their essential core at a significantly lower cost in time and resources than knowledge engineering. For example, the knowledge engineering of patient-image-retrieval heuristics at the University of Arizona's University Medical Center took eight staff-months to complete and involved multiple radiologists as well as knowledge engineers. The time required for our network learning process was only eight to 12 hours on a DECstation 3100/25 MIPS platform-several orders of magnitude shorter. With this kind of advantage, knowledge-based technology can begin to be feasible for a broader range of real-world applications. </p>
C. Wei, O. R. Liu Sheng and P. J. Hu, "Neural Net Learning for Intelligent Patient-Image Retrieval," in IEEE Intelligent Systems, vol. 13, no. , pp. 49-57, 1998.