基于SSN的数据集预处理框架

Preprocessing framework of data set based on SSN

为解决由于染色差异造成的组织病理学图片良性、恶性检测精度下降的问题,提出一种基于染色风格归一化算法的数据集预处理框架。搭建图片转换网络和损失函数网络,在训练时,联合内容图片和目标染色图片构建损失函数,通过对大量内容图片进行训练,将目标染色图片的染色风格编码在模型的权重之中;在运行时,该模型可以使输入的图片在保持原有病理学样式的情况下达到染色分布的统一。实验结果表明,经该框架处理后的数据集有着更加集中的色彩空间,使用处理后数据集训练得到的模型有着更高的分类精度。

To solve the problem of decreased accuracy of benign and malignant histopathological images detection caused by staining differences,a data set preprocessing framework based on staining style normalization algorithm was proposed.The framework included an image transfer network and a loss network.In the training stage,features were extracted from the content ima-ge and the target staining image to construct the loss function.The staining style of the target staining image was encoded in the weight of the model by training on a large number of content images.In the running stage,the model made input images achieve the uniformity of the staining distribution while maintaining the original pathological style.Experimental results show that the data set processed using this framework has a more concentrated color space,and the model trained on the processed data set has higher classification accuracy.