A Computer-Aided Diagnosis Framework for Pulmonary Nodule Interpretation in Thoracic Computed Tomography

Computer-aided Diagnosis (CAD) systems can assist radiologists in several diagnostic tasks. In this paper, we present a CAD framework which assist radiologists in interpreting pulmonary nodule in computed tomography (CT). Our approach can be considered as an approach to bridge the semantic gap between computed image features and nodule characteristics normally used by radiologists to diagnose lung cancer. The system can act as a second reader which provides automated interpretation in order to assist radiologists to interpret these characteristics more accurately and to reduce the variability in radiologists’ interpretations. In this study, our data were generated from 29 cases of thoracic CT scans collected by the Lung Image Database Consortium (LIDC). Every case was annotated by up to 4 radiologists by marking the boundaries of nodules and assigning 9 characteristics to each identified nodule. Fifty-nine image features including shape, size, gray-level intensity, and texture information were extracted from each segmented nodule image. Logistic regression was applied to generate prediction models for each characteristic based on image features. From our preliminary experimental results, we found promising mappings from computed image features to several characteristics (calcification, internal structure, malignancy, subtlety, and texture).