作者:杨晓娜，李超炜，邵慧丽，何勇军
Authors:YANG Xiao na,LI Chao wei,SHAO Hui li,HE Yong jun
摘要:目前利用深度学习识别宫颈异常细胞有两个难题:(1)宫颈细胞种类多样且宫颈细胞图像因人而异 ; (2) 宫颈细胞呈现长尾分布 ,影响宫颈细胞的分类精度 。本文提出了 一种基于深度学习的宫颈细胞分割与分类框架 。本框架首先进行细胞核分割 ,使用 U-Net 作为基础模型进行减层 ,加入 AG 模块 ,并使用 ACBlock 模块代替传统标准卷积块 ;然后使用 ResNeSt 对分割数据进行粗分类 ,将根据医生经验提取的人工特征和 ResNeSt 网络提取的机器特征进行融合进行细分类 ,利用主动学习迭代地扩充宫颈细胞类别 ,并在 BBN 模型中融合ACBlock 模块处理 长尾数据 ;最后根据TBS 诊断标准和医生的诊断经验提炼出异常细胞的诊断指标 ,筛选异常细胞 。实验表明 ,本文的分割算法较原方法提升了 3. 52% ,加入所有特征的分类算法提升了1. 2% 。针对阳性病人 ,癌细胞诊断准确率达 到 91% 。
Abstract:Currently, there are two challenges in identifying abnormal cervical cells using deep learning: (1) cervical cells are diverse and cervical cell images vary from person to person. (2 ) Cervical cells show long-tailed distribution, which affects the classification accuracy of cervical cells. In this paper, a deep learning-based cervical cell segmentation and classification framework is proposed. This framework first performs cell nucleus segmentation, uses U-Net as the base model for layer reduction, adds AG module, and uses ACBlock module instead of traditional standard convolutional blocks; then uses ResNeSt for coarse classification of segmented data, fuses manual features extracted based on physicians ′ experience and machine features extracted by ResNeSt network for fine classification , and uses active learning iteratively to expand the cervical cell categories and fuse the ACBlock module in the BBN model to process the long-tail data; finally, the diagnostic indexes of abnormal cells are refined and abnormal cells are screened according to the TBS diagnostic criteria and the physician ′s diagnostic experience. Experiments show that the segmentation algorithm in this paper improves 3. 52% compared with the original method, and the classification algorithm that incorporates all features improves 1. 2% . For positive patients, the accuracy of cancer cell diagnosis reached 91% .

PDF全文下载地址:

可免费Download/下载PDF全文