基于卷积神经网络的细胞核智能分割研究

A Nuclei Segmentation Research Based on Convolutional Neural Network

摘要: 在许多疾病的病理学诊断中，细胞核的形状、特征的变化是病变发生与否的重要依据，利用计算机智能分割出病理组织切片中的细胞核能为疾病诊断提供更多的参考。本研究将卷积神经网络应用在乳腺癌病理组织切片图像中的细胞核分割上。在对图像进行光学预处理后，将其分割成多个小图像输入到改进的Alexnet模型中对模型进行训练，使其能自动识别细胞核特征。随后，将训练后的模型用于测试集图像的细胞核分割中，把图像分割成多个小图像让模型并行处理，并最终整合所有的输出结果生成一张完整的细胞核分割图，达到细胞核分割的目的。结果表明，模型对训练集中的细胞核识别率达到92%，训练后的模型对人工标记图像中并没有标记出来的细胞核都能准确地识别出来，表明模型已成功的学习到细胞核的主要特征。最后，对测试集图片进行分割的结果显示，训练后的模型成功地把病理组织切片图像中的细胞核准确且快速地分割出来，证明这种切分图像进行细胞核分割最后再整合的方法在保证准确性的同时也能提高计算效率。

Abstract: In the pathological diagnosis of many diseases, the change of the shape and characteristics of the nucleus is an important symptom for the occurrence of the disease. Applying computer intelligence to segment the nuclei in the pathological tissue section can provide more advices for disease diagnosis. In this study, convolutional neural network was applied to the nuclei segmentation of breast cancer histopathological section image. After optical preprocessing the images, each of them was divided into multiple small images and used to train the improved AlexNet model. Then, the trained model is used in the nucleus segmentation of the test set. We divided the whole image into multiple small images, the small images were processed parallelly by the trained model, and finally integrated all the output to a whole nucleus segmentation image. The results show that the nucleus recognition rate in the training set reach to 92%. The trained model can accurately recognize all nuclei which are not labeled in the artificially labeled image, indicating that the model has success-fully learned the main features of the nucleus. Finally, the result of image segmentation in test set shows that the trained model successfully segmented the nucleus of pathological tissue slice image accurately and quickly, which proves that our method of cutting image to parallelly process and then integrating all outputs ensures both accuracy and calculation efficiency.