作者:赵司琦，梁義钦，秦健，何勇军
\n

Authors:ZHAO Si-qi,LIANG Yi-qin,QIN Jian,HE Yong-jun
摘要:摘要:利用深度学习方法的宫颈异常细胞识别通常需要大量的训练数据，然而这些数据不可避免的使用不同样本的宫颈异常细胞参与模型训练，天然地缺失了单个样本内正、异常细胞内对照，导致宫颈异常细胞的识别精度不高，假阳性率高。为解决这一问题，提出了一种基于样本基准值的宫颈异常细胞识别方法。该方法首先通过Mask R-CNN模型对宫颈细胞进行识别和宫颈细胞核分割；然后计算关键性宫颈细胞核指标，提出基准细胞概念，定义样本基准值，量化诊断标准；最后利用异常细胞核指标和模型信息完成宫颈异常细胞重分类，模拟医生对比样本内正常细胞的形态识别异常细胞，实现样本内宫颈正、异常细胞对照识别。实验表明，该方法在宫颈细胞涂片数据集上的阳性细胞查全率、阳性细胞检出准确率、样本检测准确率分别达到84.7%、94.6%、92.4%。
Abstract:Abstract:The identification of cervical abnormal cells using deep learning methods usually requires a large amount of training data, but these data inevitably use different samples of cervical abnormal cells to participate in model training, and naturally miss the positive and abnormal intracellular controls of a single sample, resulting in the fact that recognition accuracy of cervical abnormal cells is not high, and the false positive rate is high. To solve this problem, this paper proposes a method for identifying cervical abnormal cells based on sample benchmark values. Firstly, this method identifies cervical cells and segments cervical cell nucleus by Mask R-CNN model. Then we calculate the key cervical nucleus indicators, propose the concept of benchmark cells, define the sample benchmark values, and quantify the diagnostic criteria.Finally, the abnormal nucleus indicator and model information are used to complete the reclassification of abnormal cervical cells, and the abnormal cells were identified by simulating a doctor comparing the morphology of the normal cells in the sample. Experiments show that the positive cell completion rate, positive cell detection accuracy and sample detection accuracy rate on the cervical cell smear dataset reached 84.7%, 94.6% and 92.4%, respectively.

PDF全文下载地址:

可免费Download/下载PDF全文