适于多尺度宫颈癌细胞检测的改进算法北大核心CSCD

深度学习技术因其强大的特征提取能力而被广泛应用于目标检测任务中。针对多尺度宫颈癌细胞的识别准确率不均衡、检测效率低等问题,本文提出一种基于YOLO v3模型的改进识别算法mo-YOLO v3(mini-object-YOLO v3)。选用20倍数字扫描仪下采集的宫颈细胞图像作为数据集,为提高算法的鲁棒性,引入对比度增强、灰度图、旋转和翻转等多种数据增强策略扩充数据集;模型以Darknet53网络结合注意力机制作为主干模块,针对宫颈癌细胞尺寸差异大的特点,提出一种多尺度特征融合算法来优化模型结构;针对小目标检测精度低的问题,提出一种改进的损失函数,采用相对位置信息的方法减弱物体框对检测结果的影响。测试结果表明,本文所提的mo-YOLO v3模型不仅在总体识别精度上有明显的优势,同时大大提高了小尺寸宫颈癌细胞的定位精度。该模型对宫颈癌细胞识别的准确率达到90.42%,查准率达到96.20%,查全率达到93.77%,相似指数ZSI为94.97%,高于同类算法。

Improved algorithm of multi-scale cervical cancer cells detection

Deep learning technology is widely use d in target detection tasks because of its powerful feature extraction capabilities.Aiming at the problems of uneven recognition ac curacy and low detection efficiency of multi-scale cervical cancer cells,this paper proposes an improved recognition algorithm,mini-object-YOLO v3 (mo-YOLO v3) based on the YOLO v3 model.The cer vical cell images collected under a 20× digital scanner are selected as the data set.In or der to improve the robustness of the algorithm,multiple data enhancement strategies such as contra st enhancement, grayscale image,rotation and flipping are introduced to expand the data set;the model takes Darknet53 network combined with attention mechanism as the backbone module,for the large difference in the size of cervical cancer cells,a multi-scale feature fusion a lgorithm is proposed to optimize the model structure.In order to solve the problem of low detection acc uracy of small targets, an improved loss function is proposed,adopting the relative position informatio n method to reduce the influence of the object frame on the detection result.The test results show that the mo-YOLO v3 model proposed in this paper not only has obvious advantages in overall recognit ion accuracy,but also greatly improves the positioning accuracy of small-size cervical cancer cells.The model has an accuracy rate of 90.42% for identification of cervical cancer cells,a precision rate of 96.20%,a recall rate of 93.77%,and a similarity index ZSI of 94.97%,which is higher than similar algor ithms.