作者:李超炜，杨晓娜，赵司琦，何勇军
Authors:LI Chaowei,YANG Xiaona,ZHAO Siqi,HE Yongjun摘要:由于宫颈细胞样本的液基薄层细胞学检测( thin prep cytologic test，TCT) 图像内容复杂，背景颜色丰富多样，而且不同女性的宫颈细胞具有一定程度的天然差异，这给宫颈异常细胞的检测带来了很大的困难 。为解决这一难题，提出了一种名为基于特征压缩与激发和可变形卷积( SE-ResNet-deformable convolution you only look once， SER-DC YOLO) 的目标检测网络 。该网络在 YOLOv5 的 Backbone 中融合注意力机制，添加了 SE-ResNet 模块，然后改进了 SPP 层的网络结构，并且使用可变形卷积来替换普通卷积，最后修改了边界框的损失计算函数，将广义交并比( generalized intersection over union，GIoU) 改为 α-IOU Loss 。 实验表明，该网络与 YOLOv5 网络相比，在宫颈图片 数据集上召回率提高 了 19. 94% ，精度提高 了 3. 52% ，平均精度均值提高 了 7. 19% 。 相关代码链接: https: / / github. com / sleepLion99 / SER-DC_YOLO。
Abstract:Due to the complex content of Thin Prep Cytology Test ( TCT) images of cervical cell samples with rich and diverse background colors and a certain degree of natural variation of cervical cells among different women，this poses a great difficulty in the detection of abnormal cervical cells. To solve this challenge，a target detection network called SE-ResNet-Deformable Convolution You Only Look Once( SER-DC YOLO) is proposed. The network incorporates the attention mechanism in YOLOv5s Backbone，adds the SE-ResNet module ，then improves the network structure of the SPP layer and replaces the normal convolution with deformable convolution，and finally modifies the loss calculation function of the bounding box by replacing the Generalized Intersection over Union ( GIoU) to α-IOU Loss. Experiments show that the network improves recall by 19. 94% ，precision by 3. 52% ，and average precision by 7. 19% on the cervical image dataset compared with the YOLOv5 network. Link to related code: https: / /github. com / sleepLion99 / SER-DC_YOLO.


PDF全文下载地址:

可免费Download/下载PDF全文