Automated learning of patient image retrieval knowledge: Neural networks versus inductive decision trees

Retrieving a patient's prior examination images that are relevant to the current ones is a critical component in radiologists' primary examination reading services. The important role of such image retrieval support will be greatly accentuated in digital radiology practice. Radiologists' knowledge of patient prior image retrievals is rooted in their interpretation and application of the pertinent underlying medical/radiological knowledge as well as in their clinical training and experiences. At the same time, this knowledge may vary with individual practice preferences and styles, and may dynamically evolve over time. The complexity and dynamics suggest that patient image retrievals are a promising area for artificial intelligence-based automated learning techniques. Automated learning of patient image retrieval knowledge can provide continuous knowledge repository update support in an image retrieval knowledge-based system. However, the implementation of the learning techniques needs to address several challenges that include missing and noisy data, as well as multiple decision outcomes. Two techniques based on salient automated learning paradigms, neural network and symbolic learning, are investigated. Specifically, we describe the design or extension of each learning technique to address the unique characteristics of patient image retrieval knowledge and compare the resulting learning performances. The results show that the knowledge derived from the automated learning methods can achieve effective image retrievals that are comparable to those based on a knowledge-engineer-driven approach.