改进U型卷积网络的细胞核分割方法

An Improved U-Net for Cell Nuclear Segmentation

针对经典U型卷积网络在细胞核分割过程中对距离相近目标的边界较难区分、对模糊目标产生误识别等问题,提出一种改进的U型卷积网络(DU-Net)模型。为增强目标边界特征,提出一种梯度融合方法,计算样本梯度信息并将梯度图多尺度融合至U-Net编码器。解码器浅层特征通过卷积上采样密集连接至深层特征,增加特征的复用性。针对梯度消失问题,DU-Net模型在每个卷积层后采用批归一化和ReLU激活结构。针对经典U-Net模型对模糊目标的误识别问题,提出一种改进的交叉熵损失函数,该损失函数降低了模糊背景点对模型的干扰,同时提高了模型对小目标的识别能力。在2018年数据科学碗公布的670张图片、约29 500个细胞核的公开数据集上验证了DU-Net模型,结果表明,模型的预测结果与真实标签在Dice系数和Jaccard相似系数两项评价指标上分别达到95.9%和91.0%,性能优于U-Net和SegNet编码器,显著优于经典卷积神经网络模型FCN-8s。

An improved U-Net model, called DU-Net model, is proposed for the problem that the classical U-Net is difficult to distinguish the boundary of the targets which are close to each other, and often misidentifies the fuzzy targets in the process of cell nuclear segmentation. The proposed model obtains a clear target boundary feature map by calculating gradient information of samples, and then concatenates it into the U-Net encoder to enhance the edge feature of targets. This method is called gradient fusion. The shallow feature maps of the decoder are densely connected to deep features through convolutional upsampling so that the reusability of features increases and the performance of the model is improved. The DU-Net uses batch normalization and a ReLU activating function after each convolutional layer to solve gradient-vanishing problem. An improved cross entropy loss function is proposed to reduce the interference of fuzzy background points and the misidentification of small targets. The loss function improves the model’s ability to recognize small targets. Simulation results on a public dataset containing 640 images and approximately 29 500 nuclei show that the prediction results of DU-Net model reaches 95.9% and 91.0%, respectively, in the Dice coefficient and the Jaccard similarity coefficient. The performance of DU-Net is better than that of U-Net and SegNet, and is significantly better than that of the classical convolutional neural network FCN-8 s.