多流卷积神经网络细胞分类算法在宫颈脱落细胞学诊断中的价值

Multi-stream convolutional neural networks (MS-CNN) classification algorithm in cervical cytology

目的探讨人工智能在宫颈脱落细胞学诊断中的价值。方法对202例宫颈薄层液基细胞学标本采用多流卷积神经网络细胞分类算法进行判读,以人工判读结果为标准,评估其诊断效能。结果202例标本中有效筛查病例182例,多流卷积神经网络二分类和多分类算法的符合率分别为68.13%和73.08%。二分类算法的灵敏度、特异度、阴性预测值分别为100%、62.09%、100%;多分类算法的特异度相对较高,NILM、ASCUS、LSIL、ASC-H、HSIL分别为96.55%、73.75%、98.29%、99.45%、98.90%,阴性预测值为99.03%。结论人工智能的参与可提高宫颈脱落细胞学筛查效率,且多流卷积神经网络细胞分类算法在检出异常细胞方面表现突出。二分类算法漏诊率为零,故有望部分甚至完全替代细胞病理学医师的宫颈癌初筛工作,多分类算法误诊率相对较低,但存在漏诊;该算法需进一步学习完善,以达到漏诊率和诊断符合率之间更好的平衡,实现精确分类(TBS标准)的最终目标。

Objective:To evaluate the performance of MS-CNN classification algorithm.Methods:202 cases of liquid-based Pap smear were collected,and to evaluated the performance of MS-CNN in comparison with pathologists.Results:182/202 cases were successfully scanned.When applied to 2-class classification problem,MS-CNN achieved precision of 68.13%,sensitivity of 100%,specificity of 62.09%,and negative predictive value(PV-) of 100%.The results obtained from 5-class classification problem had the precision of 73.08%,and the PV- of 99.03%.Conclusions:Artificial intelligence can improve the efficiency of cervical cytology screening,and MS-CNN 2-class classification algorithm has a perfect performance in detecting abnormal cells.Improvement is still needed to reduce misdiagnosis rate,and to achieve the ultimate goal of accurate classification according to TBS standards.