全尺度上下文融合网络用于医学细胞核分割

Full Scale Context Fusion Network for Medical Nuclear Segmentation

针对细胞核图像边界模糊、对比度低,且细胞间易发生粘连等特点而导致难以准确分割的问题,本文提出了一种全尺度上下文融合网络用于细胞核的精准分割。首先,该模型结合全局上下文信息,设计了一个高级金字塔传导模块,用于对跳跃连接进行重构,为解码器提供全局信息传导流;其次,在编码器顶部创新性地加入了尺度聚合模块,该模块通过自学习可以动态地为不同尺度的目标选择合适的感受野,更好地融合多尺度上下文信息;同时,为了更好地利用最有用的特征通道,在上采样阶段加入了通道注意力机制;最后,使用改进的混合损失函数解决类不平衡的问题。在Data Science Bowl 2018和TCGA 2个数据集上进行实验,结果表明,所提出的算法能够提高对细胞核的分割性能。

A full scale context fusion network for accurate segmentation of cell nuclei is proposed,which addresses the issues of blurred boundaries,low contrast,and easy adhesion between cells in cell nucleus images.Firstly,combined with global context information,an advanced pyramid transmission module is designed to reconstruct the jump connection and provide global information transmission for the decoder.Secondly,the scale aggregation module is innovatively added at the top of the encoder.Through self-learning,the module can dynamically select appropriate receptive fields for different scale targets and better integrate multi-scale background information.Meanwhile,in order to make better use of the most useful feature channels,channel attention mechanism is added in the up-sampling stage.Finally,the improved mixed loss function is used to solve the class imbalance problem.The test results on Data Science Bowl 2018 and TCGA 2 data sets show that the proposed algorithm can improve the performance of nucleus segmentation.