Accurate breast cancer screening demands that radiologists maintain a balance of high sensitivity and high specificity when interpreting mammography. Subspecialty-trained breast imagers perform significantly better than general radiologists by recognizing more breast cancers and minimizing benign biopsies. An automated reasoning system called a Bayesian network (BN), has the potential to improve the performance of general radiologists, who interpret the majority of mammograms, to the level of subspecialty-trained breast radiologists, who are in short supply. A BN is a probabilistic graphical model that has been used for decision support in a variety of domains, including radiology. Machine learning provides an appealing way to create and optimize BNs to perform at high levels of sensitivity and specificity. Our research group has developed a prototype, expert-defined BN that uses imaging features and demographic risk factors to classify abnormalities on mammograms as benign or malignant. Our BN can perform at a higher level than general radiologists, but our goal is for it to perform as well or better than subspecialty-trained breast radiologists. To this end, we have compiled a structured, multi-relational dataset of mammography abnormalities and pathologic outcomes from which cutting edge machine learning techniques can construct an improved BN. In the past, BNs were trained and tested on individual abnormalities in isolation. In fact, other abnormalities on the same mammogram or previous mammograms, which necessarily appear in other rows of a relational database, can further improve BN learning. We have used basic statistical relational learning (SRL) techniques to enhance Bayesian learning algorithms to leverage this important additional data. A novel SRL capability introduced by our team within the last year, called view learning, makes it possible to incorporate this related data from other parts of the database by automatically defining new database fields. In our preliminary work, we have shown a stepwise improvement in BN performance: first, with conventional BN learning; then, with basic SRL; and finally, with view learning. We now seek support to determine whether more tightly integrated SRL and view learning will significantly improve our BN's ability to accurately diagnose breast cancer. In addition, we propose to investigate whether two other promising machine learning techniques, predicate-invention and collective-classification, can optimize the BN to perform at levels significantly better than current clinical practice. ? ? ?
Showing the most recent 10 out of 43 publications
