Pleural Nodule Detection in Helical CT Images Using Directional Gradient Concentration Method

Lung cancer is one of the most lethal kinds of the cancer worldwide. Computed Tomography (CT) has been shown as the most sensitive imaging modality for detecting small pulmonary nodules. To support radiologists in the challenging task of interpreting screening lung CT scans, we have developed a CAD system for the automated detection of the pleural nodules. In this study, we introduce a 3-step region of interest (ROI) hunter algorithm for the initial selection of the nodule candidates. We applied a combination of the 3D directional-gradient concentration method and the 2D morphological opening procedure on the pleural surface already segmented by an iso-surface triangulation technique called marching-cubes algorithm. Then within each ROI a nodule segmentation technique based on the morphological opening is applied to determine the 3D shape of the nodule candidate. Thirteen geometrical and textural features are extracted from the segmented nodules and then classified by a three-layered feed-forward neural network which was trained by the cross-validation method."nThis method has been validated on 102 pleural no-dules with diameter greater than 5 mm in 42 CT scan. Our CAD system achieved sensitivity values in the range of 80-85% at 35-43 FP/scan, corresponding approximately to 0.12-0.14 FP/slice and has the ability to reject more than the 93% of the FP findings generated by the ROI hunter. The average area Az under the ROC curve obtained was 0.91. The CAD system we proposed in this study has demonstrated a significant potential as a useful diagnostic aid in assisting the radiologists in the accurately identifying the pleural nodules on thoracic CT images. Current efforts are focused on further reducing the amount of false-positive findings generated by the system.