融合多尺度卷积的端到端宫颈细胞分割

End⁃to⁃end cervical cell segmentation with multi⁃scale convolution

宫颈癌是目前唯一一种病因明确的妇科恶性肿瘤,通常采用宫颈细胞筛查来进行早期排查以及治疗。在细胞筛查过程中,为了发现早期的宫颈异常细胞,需要从显微细胞图像中准确分割出细胞核与细胞质。现有宫颈细胞分割方法存在计算量大、精确率低、数据不平衡所导致的学习困难等问题。为了提高宫颈细胞分割算法的性能,引入了多尺度卷积的思想,采用一种U型编码器-解码器模型,设计了一个端到端的细胞分割算法IR U⁃Net。主要包括:①利用多尺度卷积结构加宽网络,避免人工选择卷积核,能提取多尺度特征,同时在卷积结构中加入残差连接,来减少梯度消失等现象;②通过使用Leaky⁃ReLU减少“神经元死亡”导致的网络稀疏特征多、难以收敛的问题;③采用改进的损失函数Focal⁃Dice Loss以缓解数据不平衡的问题。仿真实验结果表明,改进后的模型相比对照算法在精度上有所提高,分割性能得到改善。

Cervical cancer is the only gynecological malignant tumor with clear etiology.Cervical cell screening is usually used for early detection and treatment.In the process of cell screening,it is necessary to segment the nucleus and cytoplasm accu⁃rately from the microscopic cell images in order to find early abnormal cervical cells.The existing methods of cervical cell segmen⁃tation have some problems,such as large amount of calculation,low accuracy and learning difficulties caused by data imbalance.In this paper,the idea of multi⁃scale convolution is introduced,and an end⁃to⁃end cell segmentation algorithm IR U⁃Net is de⁃signed by using a U⁃shaped encoder⁃decoder model to improve the performance of cervical cell segmentation algorithm.It mainly includes:①Using multi⁃scale convolution structure to broaden the network,avoid manual selection of convolution kernel,and can extract multi⁃scale features.At the same time,residual connection is added to the convolution structure to reduce the phenomenon of vanishing gradient;②Leaky⁃RelU can be used to solve the problem of network has many sparse features and is difficult to con⁃verge due to“neuron death”;③Focal Dice Loss,as an improved Loss function was used to alleviate the problem of data imbalance.The simulation results show that the improved model has better accuracy and better segmentation performance than the control al⁃gorithm.