一种全新的基于胸片计算机辅助检测肺结节方案

A New Computer Aided Diagnostic Scheme for Lung Nodule Detection on Chest Radiographs

针对目前基于胸片肺结节计算机辅助检测系统的检出率低,且检测结果有大量假阳性的问题,提出一种全新检测方案.该方案首先引入基于活动形状模型的算法分割肺区,在肺区中选取大量可疑结节,然后为每个可疑结节提取基于分割结果的27个特征,最后引入线性分类器对可疑结节进行分类,给出最终检测结果.方案中,由于两步结节增强技术的引入,使得只有少量真实结节在可疑结节选取过程中丢失.特征提取时,引入分水岭算法分割可疑结节,基于分割结果提取能够有效区分可疑结节中真实结节和假结节的形状特征、灰度统计特征、曲面特征和梯度特征等,并利用可疑结节分割结果与感兴趣区域中Canny算子边缘检测结果的相关性来降低假阳性.本文选择日本放射技术学会提供的公共数据库测试系统的肺结节检测性能,系统在平均每幅图4.5个假阳性水平下检测出72.2%的结节.对非常不明显和极其不明显结节,系统的检测性能在4.5个假阳性水平下达到了52.7%.

Major challenges in current computer-aided diagnostic （CADe） schemes for nodule detection in chest radiographs （CXRs） are to improve the sensitivity and reduce the false positive rate. This paper proposed a new CADe scheme. First, an active shape model based algorithm was applied to segment the lung field. Then, seventy nodule candidates were selected and 27 features for each candidate were extracted. Linear classifier was employed for classification. Because of two-step nodule enhancement, only few nodules were missed at the candidate selection step. For feature extraction, watershed algorithm was applied to segment the nodule candidate. Based on the segmentation result, shape-based features, gray level statistics features, surface-based features and gradient features were extracted. Another feature based on Canny edge detector was designed to eliminate bone crossings as false positives.A publicly available database containing 126 nodules in 126 CXRs was used for testing our CADe scheme. A sensitivity of 72.2 % with 4.5 FPs per radiograph was achieved in a leave-one-out cross-validation test. For the very subtle and extremely subtle nodules, a sensitivity of 52.7 % was achieved at the same FPs.