Objectives: The purpose of this study was to develop and validate a computer-aided diagnosis (CAD) tool for automatic classification of pulmonary nodules seen on low-dose computed tomography into solid, part-solid, and non-solid.
Materials and methods: Study lesions were randomly selected from 2 sites participating in the Dutch-Belgian NELSON lung cancer screening trial. On the basis of the annotations made by the screening radiologists, 50 part-solid and 50 non-solid pulmonary nodules with a diameter between 5 and 30 mm were randomly selected from the 2 sites. For each unique nodule, 1 low-dose chest computed tomographic scan was randomly selected, in which the nodule was visible. In addition, 50 solid nodules in the same size range were randomly selected. A completely automatic 3-dimensional segmentation-based classification system was developed, which analyzes the pulmonary nodule, extracting intensity-, texture-, and segmentation-based features to perform a statistical classification. In addition to the nodule classification by the screening radiologists, an independent rating of all nodules by 3 experienced thoracic radiologists was performed. Performance of CAD was evaluated by comparing the agreement between CAD and human experts and among human experts using the Cohen κ statistics.
Results: Pairwise agreement for the differentiation between solid, part-solid, and non-solid nodules between CAD and each of the human experts had a κ range between 0.54 and 0.72. The interobserver agreement among the human experts was in the same range (κ range, 0.56-0.81).
Conclusions: A novel automated classification tool for pulmonary nodules achieved good agreement with the human experts, yielding κ values in the same range as the interobserver agreement. Computer-aided diagnosis may aid radiologists in selecting the appropriate workup for pulmonary nodules.