Automatic segmentation of lung cancer based on texture features of PET/CT images

Objectives It is difficult to accurately define the radiation target for radiation therapy of lung cancer. FDG-PET has been shown to be useful in helping to define such targets, but inter-observer variability in target definition remains. This study investigated whether the incorporation of texture features derived from PET/CT images could be used to automatically segment such targets.

Methods PET/CT image features derived from spatial grey-level dependence matrices, neighborhood grey tone difference matrices, Tamura's textural features, first order statistics and structural methods were investigated. A training set of images of 20 patients was used to determine the best features for discriminating between tumor and normal tissues. Using a K-nearest neighbors (KNN) classifier, the area under the receiver operating curve was calculated (AUC) and used to determine the ability of each feature to distinguish tumor from normal tissue. A decision tree was subsequently trained using KNN classifiers as nodes (DTKNN) in order to segment the tumor volume for an independent test set of images of 10 patients.

Results CT skewness and PET coarseness were found to be the most useful discriminators with AUCs of 0.705 and 0.972 respectively. Kurtosis and the standard deviation of CT morphological gradient were useful for separating heart tissue from tumor. The segmentations produced by the DTKNN on the test set were found to be quantitatively similar to contours defined by physicians.

Conclusions The method was successful in segmenting tumors using patient data taken from a different institution and scanner type than the training data. The technique was fully automated and differentiated tumor from other tissue types with similar electron densities and FDG uptake values, as it incorporated complementary information from both PET and CT.

Research Support Cancer Imaging Network of Ontario grant funded by Cancer Care Ontari