3D背景识别深度学习系统在胸腔积液检出中的价值研究

3D Context-aware Deep Learning System in Detection of Pleural Effusion:A Preliminary Study

目的探讨基于3D背景识别卷积神经网络构建的人工智能系统在胸腔积液检出中的应用价值。方法搜集2010年1月至2018年12月胸部CT平扫的检查数据2722例作为训练数集,其中含有胸腔积液的数据共2601层,所有人工智能自动识别均基于轴位图像进行标注,为了在病灶检出定位网络中考虑3D背景识别信息,选择连续7层的轴位图像作为输入。整个胸腔积液检出系统包含两个网络,病灶定位网络和肺叶分割网络,并使用3D卷积对主干网络建模。测试集选择2019年1月至2019年3月胸部CT平扫数据504例。人工智能识别系统(A组)首先检出胸腔积液是否存在,再行肺叶级别的定位,与相应病例诊断报告结果(B组)相比较。以两名有5年以上阅片经验的医师综合分析人工智能与报告结果作为金标准,进行统计学分析。结果504例CT图像中,A组检出胸腔积液347处,其中真阳性病灶175处,假阳性病灶172处。B组检出198处,其中真阳性病灶177处,假阳性病灶22处。结合人工智能和医师报告,最终发现胸腔积液218处,比最初诊断报告多发现41处。结论3D卷积神经网络的深度学习系统在胸腔积液的检出中具有应用价值。

Objective To explore the feasibility and reliability of 3 D context-aware deep learning(DL)system in detection and lobe-level localization of pleural effusion by comparison with radiologists.Methods Based on 2722 plain chest CT scanning dataset,we built Context-aware DL network for detection of pleural effusion based on deep learning,with 2601 slice-level annotations confirmed by experienced radiologists.This network took seven consecutive slices from axial CT as input and employed 3 D convolution to properly model 3 D context.Lesion position was further determained by a 3 D convolution based lobe segmentation network.Then,all the chest CT examinations were analyzed by Context-aware automatic pleural effusion(CAPE)from January to March of 2019 as group A,and the results in reporting system were classified as group B.Two experienced radiologists reviewed all the CAPE results and made evaluations as the gold standard.Results A total of 504 CT scans were tested,with 347 pleural effusion detected in group A,including 175 true positive cases and 172 false positive cases.While 198 pleural effusions were detected for group B,including 177 true positive cases and 22 false positive cases.With the CAPE results as reference,all 218 pleural effusions were detected by radiologists.Conclusion 3 D Context-aware Deep learning System demonstrated feasibility in detecting pleural effusion and localizing lung lobes.