| 陈疆红,钟朝辉,江桂莲,杨正汉,王振常,王大为.人工智能肺结节辅助诊断系统预测亚实性肺结节恶性概率[J].中国医学影像技术,2020,36(4):535~539 |
| 人工智能肺结节辅助诊断系统预测亚实性肺结节恶性概率 |
| Predicting malignant probability of subsolid nodules with artificial intelligence-assisted pulmonary nodule diagnosis system |
| 投稿时间：2019-12-12 修订日期：2020-01-22 |
| DOI：10.13929/j.issn.1003-3289.2020.04.013 |
| 中文关键词: 肺肿瘤 诊断 人工智能 体层摄影术,X线计算机 |
| 英文关键词:lung neoplasms diagnosis artificial intelligence tomography, X-ray computed |
| 基金项目:北京市"使命"人才计划项目（SML20150101）。 |
|
| 摘要点击次数: 3187 |
| 全文下载次数: 1136 |
| 中文摘要: |
| 目的 评价人工智能（AI）肺结节辅助诊断系统预测肺亚实性结节（SN）恶性概率的效能。方法 将86例接受手术治疗SN患者分为3组：组1为浸润前病变，组2为微浸润腺癌，组3为浸润性腺癌。将术前胸部CT数据导入AI肺结节识别软件，记录其自动测量的SN的CT值、体积及恶性概率预测值。比较3组SN在CT平扫、增强动脉期及延迟期中的CT值、体积及恶性概率预测值，并对各组进行平扫与增强后配对样本检验。分析根据各期CT对各组SN恶性概率预测值与CT值及体积的相关性。结果 共纳入88个SN，组1、组2和组3分别含27、28及33个SN。AI系统检测SN的敏感度为100%（88/88）。AI系统检测根据CT平扫、增强后动脉期、延迟期对组1 SN的恶性概率预测值分别为[85.18（56.64，92.08）]%、[67.15（58.99，90.30）]%和[89.82（56.64，92.23）]%，组2分别为[93.10（85.72，95.75）]%、[89.61（74.44，95.35）]%和[92.21（86.74，95.59）]%，组3分别为[97.05（92.81，98.74）]%、[96.89（90.40，98.60）]%和[96.49（89.89，98.69）]%。3期CT扫描对3组SN恶性概率预测值差异均有统计学意义（P均<0.01），且3组SN间CT值、体积差异均有统计学意义（P均<0.01）。各组平扫与增强CT恶性概率预测值比较无统计学差异（P均>0.05），各期CT对SN的恶性概率预测值与其CT值及体积均呈正相关（P均<0.01）。结论 基于深度学习的AI肺结节辅助诊断系统可协助判定肺腺癌SN侵袭程度；平扫CT数据可用于辅助预测SN恶性概率，而增强CT对判断SN性质无明显帮助。 |
| 英文摘要: |
| Objective To evaluate the efficacy of artificial intelligence (AI)-assisted pulmonary nodule diagnosis system in predicting the malignant probability of pulmonary subsolid nodule (SN). Methods Pulmonary SN from 86 patients who underwent surgical resection for pulmonary space-occupying lesions were enrolled and divided into 3 groups according to post operation pathological results, i.e. preinvasive lesions (including atypical adenomatous hyperplasia[AAH]and adenocarcinoma in situ[AIS]) in group 1, microinvasive adenocarcinoma in group 2 and invasive adenocarcinoma in group 3, respectively. Preoperative chest CT data were imported into AI pulmonary nodule diagnosis system to measure CT value and volume, also malignant probability prediction value of each SN. The differences of volume, CT value and malignant probability of SN based on plain and enhanced CT were compared among 3 groups, while the volume, CT value and malignant probability of SN were compared between plain CT and enhanced CT in each group, respectively. The correlations of the predicted malignant probability of all SN according to 3 phase CT images and nodule density and volume were analyzed, respectively.Results A total of 88 SN were enrolled, including 27 in group 1, 28 in group 2 and 33 in group 3. The sensitivity of all SN detected by AI system was 100% (88/88). The malignant probability of SN based on plain CT, arterial phase and delayed phase of enhanced CT was (85.18[56.64, 92.08])%, (67.15[58.99,90.30])% and (89.82[56.64, 92.23])% in group 1, (93.10[85.72, 95.75])%, (89.61[74.44,95.35])% and (92.21[86.74, 95.59])% in group 2, (97.05[92.81, 98.74])%, (96.89[90.40, 98.60])% and (96.49[89.89, 98.69])% in group 3, respectively. Statistical differences of nodule volume, CT value and the malignant probability of 3 phases CT images were found among 3 groups (all P<0.01), while no statistically difference of malignant probability of SN between plain and enhanced CT was found in any group (all P>0.05). The nodule CT values of arterial phase and delayed phase in each group were significantly higher than that of plain CT (all P<0.01).The predicted malignant probabilities according to plain CT, arterial phase and delayed phase enhanced CT were all positively correlated with CT value and volume of SN (all P<0.01).Conclusion The deep learning-based AI-assisted pulmonary nodule diagnosis system can assist in evaluation on the invasiveness of SN of pulmonary adenocarcinoma based on plain CT data, while enhanced CT has no significant effect on the efficiency of predicting malignant probability. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
