| 朱含笑,饶钦盼,马琳莹,等.人工智能将良性肺结节误判为高风险结节的原因分析[J].肿瘤学杂志,2024,30(1):51-56. |
| 人工智能将良性肺结节误判为高风险结节的原因分析 |
| Causes of Misdiagnosing Benign Pulmonary Nodules as High Risk Nodules by Artificial Intelligence |
| 投稿时间：2023-09-05 |
| DOI：10.11735/j.issn.1671-170X.2024.01.B009 |
|
 |
| 中文关键词: 人工智能 肺结节 误诊 计算机断层扫描 |
| 英文关键词:artificial intelligence lung nodules misdiagnosis computed tomography |
| 基金项目:浙江省中医药科技计划项目（2020ZB117）；浙江中医药大学研究所教改项目（YJSAL2022001） |
|
| 摘要点击次数: 255 |
| 全文下载次数: 150 |
| 中文摘要: |
| 摘 要： ［目的］ 探究人工智能（artificial intelligence，AI）在预测肺结节的恶性风险时将良性结节误判为高风险结节的原因。［方法］ 回顾性分析88例检查发现肺结节并于1个月内取得病理结果的患者资料，分别用AI和人工方法评估肺结节的良恶性，分析两种方法对肺结节的诊断准确率及被AI误诊的良性结节的特征。［结果］ 88例患者病理结果显示恶性结节59例，良性结节29例。AI组良性结节误诊率为82.8%（24/29），人工组为41.4%（12/29），两者对良性结节的诊断准确率差异有统计学意义（McNemar χ2＜0.001）。AI对不同大小结节组间误诊率差异有统计学意义（χ2=15.389，P＜0.001）。当良性结节出现毛刺征、分叶征、血管集束征、支气管截断征、空泡征、胸膜牵拉征等倾向于恶性结节的征象时，AI组的误诊率均大于人工组（88.2% vs 64.7%、100.0% vs 66.7%、100.0% vs 80.0%、100.0% vs 66.7%、90.0% vs 60.0%）。当出现钙化、脂肪密度倾向于良性结节的征象时，AI组的误诊率大于人工组（80.0% vs 20.0%、100.0% vs 0）。［结论］ AI对肺结节的评估存在一定的局限性，AI还需进一步完善算法，结合临床、随访、全肺整体信息，以减少误判为高风险结节的概率。 |
| 英文摘要: |
| Abstract： ［Objective］ To analyze the causes of misdiagnosing benign lung nodules as high-risk nodules by artificial intelligence (AI). ［Methods］ Imaging and pathological findings of 88 patients， who underwent biopsy or surgical treatment within 1 month after pulmonary nodules detected， were retrospectively analyzed. The pulmonary nodules on chest plain CT scan were evaluated by AI and radiologist physicians. Using pathological results as gold standard the diagnostic accuracy of both methods for pulmonary nodules was assessed， and the causes of misdiagnosing benign nodules as malignant by AI were analyzed. ［Results］ The pathological results confirmed 59 malignant nodules and 29 benign nodules. The misdiagnosis rate of benign nodules was 82.8% (24/29) in AI method and 41.4% (12/29) in the manual method (McNemar χ2 test P<0.001). When benign nodules showed spiculation sign， lobulation sign， vascular convergence sign， bronchial truncation sign， vacuole sign and pleural traction sign， the misdiagnosis rate in the AI group was higher than that in the manual group (88.2% vs 64.7%， 100.0% vs 66.7%， 100.0% vs 80.0%， 100.0% vs 66.7%， 90.0% vs 60.0%). When the nodule showing benign trending signs such as calcification and fat density， the misdiagnosis rate in the AI group was also higher than that in the manual group (80.0% vs 20.0%， 100.0% vs 0). ［Conclusion］ AI has some limitations in the evaluation of pulmonary nodules， particularly it has high misdiagnosis rate for benign nodules， indicating that AI needs to further improve its algorithm to reduce the probability of misdiagnosis. |
|
在线阅读
查看全文 查看/发表评论 下载PDF阅读器 |