人工神经网络在孤立性肺结节CT诊断研究中的应用

Application of artificial neural networks in the CT study of solitary pulmonary nodule

目的将人工神经网络理论应用于孤立性肺结节(SPN)的CT诊断研究,建立一种全新的模式判别方法,用于高分辨率CT(HRCT)或薄层CT上良恶性结节的鉴别。方法搜集经手术或穿刺病理证实的SPN 200例(原发性肺癌135例,良性结节65例),分析3项临床指标(年龄、性别及是否有痰中带血丝)和9项HRCT或薄层CT指标(部位、长径、短径、轮廓形态、毛刺、晕征、气腔密度影、结节与周围血管及胸膜的关系)。采用完全随机法从中选择140例样本作为训练集,建立人工神经网络(BP网络)诊断模型,并与软件SPSS分析处理的Logistic回归模型作比较。结果BP神经网络对所有样本的诊断符合率为98.0%(196/200),高于Logistic回归模型的符合率(86.0%,172/200)(P<0.001);ROC曲线下面积分别为0.996±0.004和0.936±0.017,差异有统计学意义(P<0.001)。结论结合神经网络理论,利用HRCT和薄层CT鉴别诊断SPN的良恶性很可能成为一种实用而可靠的临床诊断手段。

Objective To establish a new-type discriminative method in differentiating benign from malignant solitary pulmonary nodule （SPN） on high-resolution CT/thin-section CT by using artificial neural networks theory in the CT diagnostic study of SPN. Methods Two hundred SPNs pathologically proved by operation or biopsy （ primary pulmonary carcinoma 135 cases, benign nodules 65 cases） were collected, 3 clinical characteristics （age, sex, with or without bloody sputum） and 9 high-resolution CT/thin-section CT characteristics （ location, long and short diameter, contour, spiculation, halo sign, air-space, relation with the adjacent blood vessels and pleura） were analyzed. 140 cases were randomly selected to form the training samples, on which artificial neural networks model （BP networks） was built and compared with Logistic model from Statistical Package for the Social Science （SPSS） software. Results The total consistent rate of BP neural networks （98.0% , 196/200） was higher than that of Logistic model （86.0% , 172/200） （P〈0.001）. Areas under ROC curve were 0.996 ±0.004 and 0.936 ±0.017, respectively, and the difference between the two was significant （ P 〈 0. 001 ）. Conclusion Using high-resolution CT and thin-section CT in combination with artificial neural networks theory is feasible, and it is expected to become a useful and reliable clinical tool in differentiating benign from malignant SPN.