A texture feature analysis for diagnosis of pulmonary nodules using LIDC-IDRI database

This paper evaluated the performance of two-dimensional (2D) and 3D texture features from CT images on pulmonary nodules diagnosis using the large database LIDC-IDRI. Total of 905 nodules (422 malignant and 483 benign) with certain expert observer ratings of malignancy were extracted from the database based on the radiologists' painting boundaries. Feature analysis on the extracted nodules was not only based on the popular texture analysis method, e.g., the 2D Haralick texture feature model, we also explored a 3D Haralick feature model with variable directions in space. The relationships of more neighbour voxels on more directions were included for texture feature analysis. The well-established Support Vector Machine (SVM) classifier was used for the malignancy classification based on the 2D and 3D Haralick texture features. Half of the benign and malignant nodules were extracted randomly for training, and the left half nodules for testing. This operation was implemented for 100 iterations. Then the 100 classification results were shown based on the area under the curve (AUC) of the Receiver Operating Characteristics (ROC). The distinguishing results on the nodule malignancy based on the 3D Haralick texture features (Az = 0.9441) is noticeably more consistent with the expert observer ratings than that on the 2D features (Az = 0.9372).