Jean SZ Lee MRCP, MBBChir, Presenter: Nothing to Disclose

Lyndsey Clare Pickup MEng, DPhil, Abstract Co-Author: Employee, Mirada Medical Ltd

Eugene Jueren Teoh MRCP, FRCR, Abstract Co-Author: Nothing to Disclose

James Franklin MA, MBBS, Abstract Co-Author: Nothing to Disclose

Aymeric Larrue PhD, Abstract Co-Author: Employee, Mirada Medical Ltd

Mark John Gooding MENG, DPhil, Abstract Co-Author: Employee, Mirada Medical Ltd

Timor Kadir, Abstract Co-Author: Employee, Mirada Medical Ltd

Fergus Vincent Gleeson MBBS, Abstract Co-Author: Alliance Medical Ltd Consultant

Differentiating benign from malignant pulmonary nodules is critical in the management of patients with pulmonary nodules. The purpose of this study was to investigate the use of textural and imaging features to differentiate pulmonary nodules using machine learning methods.

33 patients with histology-proven pulmonary nodules were included. All patients underwent a volumetric chest CT (VCT) scan, with first dynamic contrast-enhanced chest CT (dceCT) and PET-CT scans. 23 (71.9%) were malignant. Nodules were manually contoured on the VCT and baseline dceCT scans, and propagated to the remaining scans using deformable image registration (Mirada XD, Mirada Medical, Oxford, UK). Imaging measures, such as maximum and mean intensity, and textural features, such as kurtosis or fractal dimension, were calculated considering both the full-nodule volumes and sub-volumes inside and outside the drawn contours. Volume doubling time (VDT) and SUV/TLG statistics for PET were also included to create large feature vectors with several hundred entries per nodule. Gaussian distributions were fitted to subsets of 2-3 features for the malignant and benign training populations separately. A leave-one-out paradigm was adopted (train on all-but-one datapoints; test on the withheld one). Each test nodule was classified as belonging to whichever population gave a higher likelihood score given its feature vector.

32/33 (97%) nodules were correctly classified as cancer/benign under the leave-one-out paradigm. The 3 optimal features were a “fractalness” measure on the nodule at 2 minutes post-contrast, the minimum intensity within the nodule at 4 minutes post-contrast, and a skewness measure on the core of the nodule (defined as areas not within a small distance of the contour boundary) also at 2 minutes post-contrast. These features remained optimally discriminative when the nodule dataset was entirely re-contoured by an independent researcher.

Textural analysis and imaging features using machine learning methods can help differentiate benign from malignant pulmonary nodules and help guide management.

Differentiation of benign and malignant pulmonary nodules is a common clinical problem that may be helped using textural analysis and imaging features.

Lee, J, Pickup, L, Teoh, E, Franklin, J, Larrue, A, Gooding, M, Kadir, T, Gleeson, F, Textural Analysis and Imaging Features to Differentiate Benign from Malignant Pulmonary Nodules.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14019503.html