人工智能识别技术在T1期肺癌诊断中的临床应用研究

Clinical Application of Artificial Intelligence Recognition Technology in the Diagnosis of Stage T1 Lung Cancer

背景与目的肺癌是目前国内外发病率及致死率最高的癌症,使用计算机断层扫描(computed tomography, CT)筛查肺癌结节工作量巨大。通过人工智能深度学习,在1 mm及5 mm层厚的胸部CT中,利用计算机人工智能自动寻找肺癌结节,以测试人工智能在肺癌自动识别中的效果。方法分别标注5 mm及1 mm层厚的T1期肺癌患者胸部CT片各5,000例,进行计算机神经网络学习,形成肺部结节的算法,利用人工智能形成的算法测试1 mm及5 mm层厚的T1期肺癌患者胸部CT片各500例,同人类读片进行比较,测试敏感性及特异性。结果利用人工智能读取5 mm的胸部CT 500例,敏感度达95.20%,特异性达93.20%,两次重复读取的Kappa值达0.926,1。对于1 mm的胸部CT 500例测试,敏感性为96.40%,特异性为95.60%,两次重复读取的Kappa值为0.938,6。而与5位医师相比,对1mm层厚的相同验证集CT片进行读片,人工智能与人工读片对于肺癌结节和阴性对照读片的检测率相似,两者之间比较无显著差异。而在5 mm层厚的相同验证集CT片比较中,人工智能对肺癌结节的检出数优于人工读片,敏感性更高,但误报数增多,特异性稍差。结论通过人工智能自动学习早期肺癌胸部CT图像,可以达到较高的早期肺癌识别的敏感性及特异性,可辅助医生进行诊断工作。

Background and objective Lung cancer is the cancer with the highest morbidity and mortality at home and abroad at present. Using computed tomography(CT) to screen lung cancer nodules is a huge workload. To test the effect of artificial intelligence in automatic identification of lung cancer by using artificial intelligence to find the lung cancer nodules automatically in the chest CT of 1 mm and 5 mm thick. Methods 5,000 cases of T1 stage lung cancer patients with 1 mm and 5 mm layer thickness were respectively labeled and learned by computer neural network, the algorithm of forming pulmonary nodules was carried out. 500 cases of chest CT in T1 stage lung cancer patients with 1 mm and 5 mm thickness were tested by artificial intelligence formation, and the sensitivity and specificity were compared with artificial reading. Results Using artificial intelligence to read chest CT 500 in 5 mm, the sensitivity was 95.20%, the specificity was 93.20%, and the Kappa value of two times repeated read was 0.926,1. For 1 mm chest CT 500 cases, the sensitivity is 96.40%, the specificity is 95.60%, and the Kappa reads two times is 0.938,6. Compared with 5 doctors, the same CT sets with 1 mm thickness were read. The detection rates of artificial intelligence and artificial reading were similar to those of lung cancer nodules and negative control read films, and there was no significant difference between them. In the comparison of the same CT slices with 5 mm thickness, the number of detection of lung cancer nodules by artificial intelligence is better than that of artificial reading, and the sensitivity is higher, but the number of false messages is increased and the specificity is slightly worse. Conclusion The automatic learning of early lung cancer chest CT images by artificial intelligence can achieve high sensitivity and specificity of early lung cancer identification, and assist doctors in the diagnosis of lung cancer.