改进EfficientNet的多视图特征融合的CIN诊断  

Multi-view Feature Fusion Based on Improved EfficientNet for CIN Diagnosis

作　　者：郭颖 王永雄[1] 杨慧敏 张佳鹏 孙青 GUO Ying;WANG Yongxiong;YANG Huimin;ZHANG Jiapeng;SUN Qing(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China;The First Affiliated Hospital of Wannan Medical College,Wuhu 241000,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093 [2]皖南医学院第一附属医院,安徽芜湖241000

出　　处：《控制工程》2024年第3期560-566,共7页Control Engineering of China

基　　金：上海市自然科学基金资助项目(22ZR1443700)。

摘　　要：现有的计算机辅助宫颈图像诊断方法大多是基于醋酸或Lugol’s碘的单视图图像,忽略了容易因非病理组织引起的假阳性反应而造成误诊的问题。因此,该研究将3种视角的阴道镜图像通过通道级联方式作为总输入,构建以EfficientNet为主干网络的深度学习框架,用于宫颈上皮内瘤变的诊断。此外,由于阴道镜三视图内容的高度相关性和空间一致性,需要从通道和空间2个维度进行特征加权。因此,在EfficientNet内部嵌入了卷积块注意力模块(convolutional block attention module, CBAM)的特征融合层,进一步加强病灶特征选择能力。在真实临床阴道镜数据集上进行实验,准确率和F1-Score分别达到了88.5%和88.2%,曲线下面积(area under the curve, AUC)值为0.90。实验结果表明,所提方法可以帮助临床医生进行快速的宫颈上皮内瘤变(cervical intraepithelial neoplasia, CIN)诊断,并优于已知的相关工作。Most of the existing computer-aided cervical image diagnostic methods are based on single-view images of acetic acid or Lugol’s iodine,ignoring the problem that false-positive reactions caused by non-pathological tissues are easy to be misdiagnosed.Three-view colposcopic images are used as total input by channel cascading to construct a deep learning framework with EfficientNet as the backbone network for the diagnosis of cervical intraepithelial neoplasia.In addition,due to the high correlation and spatial consistency of the contents of the three-view colposcopy,feature weighting needs to be carried out from the channel and space dimensions,so a feature fusion layer of the convolutional block attention module(CBAM)is embedded inside EfficientNet to further enhance the ability of lesion feature selection.The accuracy,F1-Score of the experiment on the real clinical colposcopy dataset reach 88.5%,and 88.2%,respectively,and the area under the curve(AUC)value is 0.90.The experimental results show that the proposed method can help clinicians to perform rapid diagnosis of cervical intraepithelial neoplasia(CIN)and outperforms known related work.

关 键 词：宫颈上皮内瘤变 阴道镜图像 深度学习 EfficientNet 多视图融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]