Automatic Classification of Perifissural Pulmonary Nodules in Thoracic CT Images

F. Ciompi, B. de Hoop, C. Jacobs, M. Prokop, P. a de Jong and B. van Ginneken

Annual Meeting of the Radiological Society of North America 2014.

Title. Automatic Classification of Perifissural Pulmonary Nodules in Thoracic CT Images. Purpose. Up to one third of pulmonary nodules detected in heavy smokers are perifissural nodules (PFNs) that do not require follow-up. An automatic method is presented to distinguish PFNs from solid nodules. Materials and Methods. We used all baseline scans with a pulmonary nodule from one of the sites of the NELSON trial. All participants were either current or former heavy smokers (age between 50 and 75 years), and underwent low-dose CT (Mx8000 IDT 16; Philips Medical Systems, Cleveland, Ohio). Human experts annotated non-calcified solid nodules in 1,729 scans, and classified these as PFN (788) and non-PFN (3,038). We formulated PFN classification as a machine learning problem where a classifier is trained to automatically label nodules as PFN or non-PFN. Given the characteristic triangular-like shape of PFNs, a novel descriptor encoding information on nodule morphology was designed. The descriptor is based on frequency analysis of intensity profiles sampled in the CT image. Given a detected nodule, spherical surfaces up to a maximum radius R are considered, centered on the center of mass of the nodule. For each sphere, the image intensity is sampled along C circular profiles on the surface of each sphere at constant angular distance. The profiles are interpreted as a periodic signal, and their spectrum is obtained using a Fast Fourier Transform. Each spectrum encodes information on nodule morphology through a set of characteristic frequencies. A set of K spectral signatures is computed applying K-means on the collected set of spectra. A compact nodule descriptor is obtained as the histogram of spectral signatures along the spheres. A Random Forests classifier with 100 trees was used for supervised learning. A 10-folds cross-validation scheme was applied to evaluate the method on the 3,826 nodules, using C=128, K=100. Since the range of PFNs diameters is 2.8-10.6 mm, we used R = 7.5 mm. Results. We obtained a value of area under the ROC curve of 0.85, with an optimal operating point of 77% sensitivity and 79% specificity. Misclassified PFNs were often close to the pleura or to other vascular structures. Conclusion. Classification of pulmonary nodules as PFN is feasible and has the potential to be used as an automatic tool in CAD. Clinical Relevance. PFNs rarely turn out to be malignant, even though their growth rate is similar to that of malignant nodules. Automatic recognition of PFNs could reduce the number of unnecessary follow-up CT exams.