A Quantitative Analysis System of Pulmonary Nodules CT Image for Lung Cancer Risk Classification

To improve the classification performance for lung nodule, we proposed a lung nodule CT image feature extraction method. The approach designed multi-directional distribution features to represent nodules in different risk stages effectively. First, the reference map is constructed using integral image, and then K-Means approach is performed to clustering the reference map and calculate its label map. The density distribution map of lung nodule image was generated after calculate the gray scale density distribution level for each pixel. An exponential function was designed to weighting the angular histogram for each components of the distribution map. Then, quantitative measurement was performed by Random Forest classifier. The evaluation dataset is the lung CT database which provided by Shanghai Zhongshan Hospital (ZSDB), the nodule risk categories were AAH, AIS, MIA, and IA. In the result the AUCs are 0.9771, 0.9917, 0.9590, 0.9971, and accuracy are 0.7478, 0.9167, 0.7450, 0.9567 for AAH, AIS, MIA and IA respectively. The experiments show that the proposed method performs well and is effective to improve the classification performance of pulmonary nodules.