基于宫颈上皮与血管特征的阴道镜图像深度学习模型探索

Deep learning model exploration of colposcopy image based on cervical epithelial and vascular features

目的通过识别阴道镜图像中的上皮与血管特征,探讨深度学习的目标检测技术在宫颈癌前病变定位及分类中的可行性。方法收集2018年3月至2019年7月复旦大学附属妇产科医院病理诊断为宫颈低级别(5708例)、高级别(2206例)癌前病变和宫颈癌(514例)患者的28975张阴道镜图像。依照国际宫颈病理与阴道镜联盟及美国阴道镜与病理协会阴道镜标准化术语,基于16类宫颈上皮与血管征象,对图像进行像素级标注后得到有效标签39858个。为降低细粒度标注可能存在的误差,进一步将标签归并为低级别、高级别和癌三大类。采用经过二次迁移学习的ResNet101预训练网络作为特征提取器,分别构建基于Faster-RCNN网络结构的高级别病变目标检测和低、高、癌三类目标检测模型。结果基于ImageNet预训练的ResNet101模型,通过宫颈转化区分类的开源阴道镜数据进行第一次迁移学习,再以自有数据的病变分类为目标进行第二次迁移学习得到特征提取器。所构建的高级别和三类病变定位检测模型在测试集上的识别精度均值mAP@IOU=0.5分别为0.82和0.67。结论利用国内最大阴道镜中心的大样本数据,基于上皮与血管特征的精细标注,深度学习模型在宫颈癌前病变检测中取得较好效果。深度学习的目标检测技术在宫颈癌前病变定位及分类中可行,尽管在识别精度上仍有提升空间,但已显示其辅助宫颈癌筛查尤其是指导定位的可行性。

Objective To explore the feasibility of object detection and deep learning model applied on the localization and classification of cervical precancerous lesions based on identifying the epithelial and vascular features in colposcopy images.Methods A total of 28975 colposcopic images were collected from Mar 2018 to Jul 2019 in the Obstetrics and Gynecology Hospital of Fudan University,including cervical low-grade lesion(5708 patients),high-grade lesion(2206 patients)and cervical cancer(514 patients).According to the colposcopy standardized terminology of the International Federation for Cervical Pathology and Colposcopy and American Society for Colposcopy and Cervical Pathology,39858 valid labels were obtained after pixel-level labeling based on 16 types of cervical epithelial and vascular signs.In order to reduce the error of fine-grained labeling,labels were further classified into three categories:low-grade,high-grade and cancer.Using ResNet101 pre-training network after secondary transfer learning as feature extractor,the models of high-grade lesion object detection and three categories(low-grade,high-grade and cancer)object detection based on Faster-RCNN network structure were constructed respectively.Results Based on the ResNet101 model of ImageNet pre-training,the first transfer learning was performed through the open source colposcopy data of cervical transformation zone classification,and the second transfer learning was carried out based on our own data of lesion classification to obtain the feature extractor.The average recognition accuracy mAP@IOU=0.5 obtained on the high-grade lesion model and the three categories model were 0.82 and 0.67,respectively.Conclusion Using the large sample data of the largest colposcopy center in China,the deep learning model can make a good performance in the detection of cervical precancerous lesions based on the fine labeling of epithelial and vascular features.Although there was still room for improvement on accuracy,these models were shown to be potential for cervical cancer screening,especially on guidance for location.