Exploring the potential of collaborative filtering for user-adaptive mammography education

Specialized training in breast imaging is critical to ensure high diagnostic accuracy of the radiologists who read screening mammograms in their daily practice. Previously, we proposed a framework for an individualized computer-aided mammography training system as a time-efficient and effective support for radiology education. The system utilizes the concept of user modeling to adapt the training protocol to meet the individual needs of the radiologists-in-training. User models are derived to predict the difficulty that a previously unseen case will pose to the modeled user. Constructing accurate models of the trainees is crucial for the overall effectiveness of the proposed training. In this paper we explore the potential of collaborative filtering for this task. Collaborative filtering is based on the assumption that the relation between ratings of different users or between ratings of different items observed for previous items will translate to new items and users. In the context of radiology trainee modeling we use this approach to predict errors that the trainees will make for unseen cases. These predicted errors can serve as the basis to identify challenging cases that are expected to be more beneficial when included in the training of a given trainee. We performed an experimental evaluation of the algorithm using data collected at Duke University Medical Center from 10 radiology residents for the problem of determining the malignancy status of masses based on their mammographic appearance. Our experiments showed that the collaborative filtering algorithm is able to distinguish cases of high and low difficulty and therefore demonstrated the promise of this approach in building adaptive computer-aided educational systems in radiology education.