一种改进U型神经网络的医学细胞核图像分割方法

Improved U-type Neural Network Method for Medical Nuclear Image Segmentation

在细胞核分割任务中,存在细胞核的边缘轮廓信息模糊以及细胞核和背景的对比度低造成难以分割的问题.针对此问题,常规的U-Net模型通过跳跃连接在降采样和上采样过程中结合高低层次的信息,具有较好的分割效果.但是,在细胞核边界的分割上仍旧存在着一定程度的过分割、欠分割等缺陷.由此本文提出一种改进的U-Net网络模型.首先,该模型采用深度特征聚合结构和高级监督的学习方法,巧妙融合不同层级的信息,达到对目标的精准分割;其次在其架构上创新性的加入层次交融模块,该模块学习各个不同层次的重要性,将学到的权重加载到分割图上;同时在嵌套的卷积块中加入注意力机制,抑制冗余特征,使得细胞核和背景更好的分割开来;最后使用改进的混合损失函数解决类不平衡的问题.在dsb2018数据集上进行测试,本方法得到Dice系数为0.8719,交并比达到0.8853.结果表明本方法能够对细胞核进行更好的分割.

In the nuclear segmentation task,the edge information of the nucleus and the low contrast between the nucleus and the background make it difficult to segment.The conventional u-net model combines high-level and low-level information in the process of down sampling and up sampling through jump connection,which makes u-net have a good segmentation effect.However,there are still some defects such as over segmentation and under segmentation in the nuclear boundary segmentation.In this paper,an improved u-net network model is proposed.Firstly,the deep feature aggregation structure is adopted in the model,and the advanced supervised learning method is adopted to overlay the deep detail information and the shallow spatial information to realize the accurate segmentation of the nucleus;Secondly,the hierarchical blending module is added innovatively in its architecture.The module learns the importance of different levels,loads the learned weights into the segmentation graph,and adds attention mechanism in the nested convolution block to suppress redundant features,so that the nucleus and background can be better separated;Finally,the improved mixed loss function is used to solve the class imbalance problem.The test results on dsb2018 data set show that the Dice coefficient is 0.8719 and the intersection union ratio is 0.8853.The results show that this method can segment the nucleus better.