在乳腺癌诊断中，特定区域内有丝分裂细胞数量是一个重要的衡量标准，它能够反映肿瘤扩散的程度，对于预测癌症的侵袭性具有重要影响。有丝分裂计数是一种耗时且具有挑战性的技术，病理学家需要通过显微镜检查用Hematoxylin和Eosin（H＆E）染色的活检切片进行手动操作。由于数据集有限以及有丝分裂细胞与非有丝分裂细胞之间的相似性，因此在H＆E染色切片中检测有丝分裂是困难的。通过协助筛选、识别和标记有丝分裂细胞，计算机辅助有丝分裂检测技术使整个过程更加容易。对于较小数据集的计算机辅助检测方法，预先训练的卷积神经网络被广泛采用。本研究调查了一个多CNN框架（使用三个预先训练的CNN）对有丝分裂检测的有效性。采用VGG16、ResNet50和DenseNet201预先训练的网络，从组织病理学数据中收集特征和识别。该提出的框架利用MITOS-ATYPIA竞赛2014提供的所有MITOS数据集的训练文件夹以及TUPAC16数据集的所有73个文件夹。每个预先训练的卷积神经网络模型，例如VGG16、ResNet50和DenseNet201，分别提供83.22%、73.67%和81.75%的准确性。不同的这些预先训练CNN的组合构成了一个多CNN框架。由三个预先训练的CNN组成的多CNN的性能指标，加上线性支持向量机，准确率为93.81%，F1值为92.41%，而与Adaboost和随机森林等其他分类器的多CNN组合相比，性能更优。版权所有© 2023 Elsevier Ltd.

In breast cancer diagnosis, the number of mitotic cells in a specific area is an important measure. It indicates how far the tumour has spread, which has consequences for forecasting the aggressiveness of cancer. Mitosis counting is a time-consuming and challenging technique that a pathologist does manually by examining Hematoxylin and Eosin (H&E) stained biopsy slices under a microscope. Due to limited datasets and the resemblance between mitotic and non-mitotic cells, detecting mitosis in H&E stained slices is difficult. By assisting in the screening, identifying, and labelling of mitotic cells, computer-aided mitosis detection technologies make the entire procedure much easier. For computer-aided detection approaches of smaller datasets, pre-trained convolutional neural networks are extensively employed. The usefulness of a multi CNN framework with three pre-trained CNNs is investigated in this research for mitosis detection. Features were collected from histopathology data and identified using VGG16, ResNet50, and DenseNet201 pre-trained networks. The proposed framework utilises all training folders of the MITOS dataset provided for the MITOS-ATYPIA contest 2014 and all the 73 folders of the TUPAC16 dataset. Each pre-trained Convolutional Neural Network model, such as VGG16, ResNet50 and DenseNet201, provides an accuracy of 83.22%, 73.67%, and 81.75%, respectively. Different combinations of these pre-trained CNNs constitute a multi CNN framework. Performance measures of multi CNN consisting of 3 pre-trained CNNs with Linear SVM give 93.81% precision and 92.41% F1-score compared to multi CNN combinations with other classifiers such as Adaboost and Random Forest.Copyright © 2023 Elsevier Ltd. All rights reserved.