Liu Yiqing, Li Xi, Zheng Aiping, Zhu Xihan, Liu Shuting, Hu Mengying, Luo Qianjiang, Liao Huina, Liu Mubiao, He Yonghong, Chen Yupeng
Department of Life and Health, Tsinghua Shenzhen International Graduate School, Shenzhen, China.
Department of Gastroenterology, Peking University Shenzhen Hospital, Shenzhen, China.
Front Mol Biosci. 2020 Aug 4;7:183. doi: 10.3389/fmolb.2020.00183. eCollection 2020.
To obtain molecular information in slides directly from H&E staining slides, which apparently display morphological information, to show that some differences in molecular level have already encoded in morphology.
In this paper, we selected Ki-67-expression as the representative of molecular information. We proposed a method that can predict Ki-67 positive cells directly from H&E stained slides by a deep convolutional network model. To train this model, we constructed a dataset containing Ki-67 negative or positive cell images and background images. These images were all extracted from H&E stained WSIs and the Ki-67 expression was acquired from the corresponding IHC stained WSIs. The trained model was evaluated both on classification performance and the ability to quantify Ki-67 expression in H&E stained images.
The model achieved an average accuracy of 0.9371 in discrimination of Ki-67 negative cell images, positive cell images and background images. As for evaluation of quantification performance, the correlation coefficient between the quantification results of H&E stained images predicted by our model and that of IHC stained images obtained by color channel filtering is 0.80.
Our study indicates that the deep learning model has a good performance both on prediction of Ki-67 positive cells and quantification of Ki-67 expression in cancer samples stained by H&E. More generally, this study shows that deep learning is a powerful tool in exploring the relationship between morphological information and molecular information.
The main program is available at https://github.com/liuyiqing2018/predict_Ki-67_from_HE.
直接从看似仅显示形态学信息的苏木精-伊红（H&E）染色切片中获取分子信息，以表明分子水平的某些差异已编码在形态学中。
在本文中，我们选择Ki-67表达作为分子信息的代表。我们提出了一种方法，可通过深度卷积网络模型直接从H&E染色切片预测Ki-67阳性细胞。为训练该模型，我们构建了一个包含Ki-67阴性或阳性细胞图像及背景图像的数据集。这些图像均从H&E染色的全切片图像（WSIs）中提取，且Ki-67表达是从相应的免疫组化（IHC）染色WSIs中获取的。对训练好的模型进行了分类性能以及量化H&E染色图像中Ki-67表达能力的评估。
该模型在区分Ki-67阴性细胞图像、阳性细胞图像和背景图像时，平均准确率达到0.9371。在量化性能评估方面，我们的模型预测的H&E染色图像量化结果与通过颜色通道滤波获得的IHC染色图像量化结果之间的相关系数为0.80。
我们的研究表明，深度学习模型在预测Ki-67阳性细胞以及量化H&E染色癌症样本中的Ki-67表达方面均具有良好性能。更普遍地说，本研究表明深度学习是探索形态学信息与分子信息之间关系的有力工具。
主要程序可在https://github.com/liuyiqing2018/predict_Ki-67_from_HE获取。
