基于非放大内镜白光图像的早期结直肠癌计算机辅助诊断(附视频)
本研究利用非放大内镜白光图像开发计算机辅助诊断(CADx)系统,用于区分早期结直肠癌的深层黏膜下浸润(T1b)与浅层浸润(T1a)或无浸润(Tis)。从10家学术医院的1470例患者中收集1513个病变(Tis 1074个;T1a 145个;T1b 294个),按3:1分为训练集和测试集,采用ResNet-50网络提取特征。在测试集(403个结直肠癌,1392张图像)中,CADx在90% cutoff下特异性为94.4%(95%CI 91.3-96.6),敏感度为59.8%(95%CI 48.3-70.4),准确率为87.3%(95%CI 83.7-90.4);CADx的AUC为85.1%(95%CI 79.9-90.4),与专家1的88.2%(95%CI 83.7-92.8)和专家2的85.9%(95%CI 80.9-90.9)相当,优于培训者1的77.0%(95%CI 71.5-82.4;vs CADx,P=0.0076)和培训者2的66.2%(95%CI 60.6-71.9;P<0.0001)。结论认为CADx系统对T1b病变诊断具有出色的每病变特异性和准确性,相当于专家水平,优于培训者。
要点
- CADx在90% cutoff下诊断T1b的每病变特异性94.4%、准确率87.3%、敏感度59.8%。
- CADx的AUC为85.1%,与专家1(88.2%)和专家2(85.9%)相当。
- CADx的AUC优于培训者1(77.0%,P=0.0076)和培训者2(66.2%,P<0.0001)。
核心结论
- 基于内镜白光图像开发的CADx系统对T1b病变诊断表现出色的每病变特异性和准确性,相当于专家水平,优于培训者。
研究设计与样本
- 从10家学术医院的1470例患者中收集1513个病变(Tis 1074个;T1a 145个;T1b 294个),按3:1分为训练集和测试集。
- 使用ResNet-50网络作为主干提取图像特征,采用过采样和焦点损失补偿浸润分期的类别不平衡。
- 测试集包括403个结直肠癌的1392张图像,由2名专家和2名培训者阅读同一测试集。
主要发现
- 在90% cutoff的每病变评分下,CADx特异性为94.4%(95%CI 91.3-96.6),敏感度为59.8%(95%CI 48.3-70.4),准确率为87.3%(95%CI 83.7-90.4)。
- CADx的AUC为85.1%(95%CI 79.9-90.4),专家1为88.2%(95%CI 83.7-92.8),专家2为85.9%(95%CI 80.9-90.9)。
- 培训者1的AUC为77.0%(95%CI 71.5-82.4;vs CADx,P=0.0076),培训者2为66.2%(95%CI 60.6-71.9;P<0.0001)。
- 该功能也在9段短视频上得到确认。
关键数值与推荐
- CADx在90% cutoff下诊断T1b的每病变特异性94.4%(95%CI 91.3-96.6)、敏感度59.8%(95%CI 48.3-70.4)、准确率87.3%(95%CI 83.7-90.4)。
- CADx的AUC为85.1%(95%CI 79.9-90.4),专家1为88.2%(95%CI 83.7-92.8),专家2为85.9%(95%CI 80.9-90.9)。
- 培训者1的AUC为77.0%(95%CI 71.5-82.4;vs CADx,P=0.0076),培训者2为66.2%(95%CI 60.6-71.9;P<0.0001)。
临床怎么用
- 原文未据此给出临床推荐。
局限性
- 原文未报告局限性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
BACKGROUND AND AIMS: Differentiation of colorectal cancers (CRCs) with deep submucosal invasion (T1b) from CRCs with superficial invasion (T1a) or no invasion (Tis) is not straightforward. This study aimed to develop a computer-aided diagnosis (CADx) system to establish the diagnosis of early-stage cancers using nonmagnified endoscopic white-light images alone. METHODS: From 5108 images, 1513 lesions (Tis, 1074; T1a, 145; T1b, 294) were collected from 1470 patients at 10 academic hospitals and assigned to training and testing datasets (3:1). The ResNet-50 network was used as the backbone to extract features from images. Oversampling and focal loss were used to compensate class imbalance of the invasive stage. Diagnostic performance was assessed using the testing dataset including 403 CRCs with 1392 images. Two experts and 2 trainees read the identical testing dataset. RESULTS: At a 90% cutoff for the per-lesion score, CADx showed the highest specificity of 94.4% (95% confidence interval [CI], 91.3-96.6), with 59.8% (95% CI, 48.3-70.4) sensitivity and 87.3% (95% CI, 83.7-90.4) accuracy. The area under the characteristic curve was 85.1% (95% CI, 79.9-90.4) for CADx, 88.2% (95% CI, 83.7-92.8) for expert 1, 85.9% (95% CI, 80.9-90.9) for expert 2, 77.0% (95% CI, 71.5-82.4) for trainee 1 (vs CADx; P = .0076), and 66.2% (95% CI, 60.6-71.9) for trainee 2 (P < .0001). The function was also confirmed on 9 short videos. CONCLUSIONS: A CADx system developed with endoscopic white-light images showed excellent per-lesion specificity and accuracy for T1b lesion diagnosis, equivalent to experts and superior to trainees. (Clinical trial registration number: UMIN000037053.).
引用
- Nemoto D, Guo Z, Katsuki S, et al. Computer-aided diagnosis of early-stage colorectal cancer using nonmagnified endoscopic white-light images (with videos). Gastrointest Endosc. 2023. PMID: 36738793. doi:10.1016/j.gie.2023.01.050
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。