A Strategy for SPN Detection Based on Biomimetic Pattern Recognition and Knowledge-Based Features

摘要

Image processing techniques have proved to be effective in improving the diagnosis of lung nodules. In this paper, we present a strategy for solitary pulmonary nodules (SPN) detection using radiology knowledge-based feature extraction scheme and biomimetic pattern recognition (BPR). The proposed feature extraction scheme intends to synthesize comprehensive information of SPN according to radiology knowledge, e.g. grey level features, morphological, texture and spatial context features. Using support vector machine (SVM), Naive Bayes (NB) and BPR as the classifiers to evaluate different feature representation schemes, our experimental study shows that the proposed radiology knowledge-based features can significantly improve the classification effectiveness of SPN detection from nonnodules, in terms of accuracy and F_1 value, regardless of the classifiers used. We also note that BPR can deliver a consistent performance using our knowledge-based features, even the ratios between nonnodules and nodules are quite different in the training set.