Keywords:
Cancer, Computer Applications-General, Computer Applications-Detection, diagnosis, Computer Applications-3D, CT, Lung, Computer applications
Authors:
R. Casale1, B. M. Maris2, C. Casale3, E. Bertelli4, D. Caramella5, P. Fiorini2; 1Portogruaro (VE)/IT, 2Verona/IT, 3Rome/IT, 4Florence/IT, 5Pisa/IT
DOI:
10.1594/ecr2018/C-1300
Methods and materials
We have evaluated 65 CT scans of lungs.
Dataset were randomly extracted from the SPIE-AAPM Lung CT Challenge archive [5,
6,
7,
8].
Studies were obtained on multi-detector CT system (Philips Brilliance).
All CT scans were performed with high resolution convolution kernel and 1 mm reconstruction slice thickness; 18 CT scans were unenhanced,
whereas 47 CT scans were acquired after iodine contrast administration.
All nodules within these scans were classified by an expert radiologist to be either primary lung cancer or benign nodules,
based on follow-up imaging and/or pathologic assessment [7].
Pulmonary nodules were manually or semi-automatically segmented by a radiologist (RC) using 3D-Slicer 4.6.2 [9] (Figure 2).
We extracted five shape features (Volume,
Surface area,
Surface/Volume ratio,
Spherical disproportion and Sphericity) with HeterogeneityCAD module in 3D-Slicer [10] (Figure 3).
Median and interquartile range (IQR) were calculated for all the shape features.
Mann-Whitney U test was used to discriminate between the shape values of benign and malignant nodules.
Combination of five shape features was entered in a logistic model in Weka software 3.8.1 [11] with leave-one-out cross-validation testing mode.
Areas under the curves (AUC) were calculated with benign/malignant outcome.