Predictive Data Mining for Lung Nodule Interpretation

Diagnostic decision-making in pulmonary medical imaging has been improved by computer-aided diagnosis (CAD) systems, serving as second readers to detect suspicious nodules for diagnosis by a radiologist. Though increasing accurate, these CAD systems rarely offer useful descriptions of the suspected nodule or their decision criteria, mainly due to lack of nodule data. In this paper, we present a framework for mapping image features to radiologist-defined diagnostic criteria based on the newly available data from the Lung Image Database Consortium (LIDC). Using data mining, we found promising mappings to clinically relevant, human-interpretable nodule characteristics such as malignancy, margin, spiculation, subtlety, and texture. Bridging the semantic gap between computed image features and radiologist defined diagnostic criteria allows CAD systems to offer not only a second opinion but also decision-support criteria usable by radiologists. Presenting transparent decisions will improve the clinical acceptance of CAD.