PCNN深度模型及乳腺病理图像自动分析方法研究

The count of mitotic cells is a critical factor in breast cancer grading systems. Extracting the mitotic cell from the histopathological image is a very challenging task. The phenomenon of misdiagnosis and missed diagnosis often happen in the process of manual mitosis detection due to a heavy workload. Besides, the inter-observer and intro-observer variations also frequently appear regarding to the same histopathological images. Therefore, it has been an urgent need to employ computer automatic analysis technology for the quantitative analysis and assessment for the breast pathology image. .However, the accurate rates of the current mitosis detection methods are generally low, which is far from meeting the clinical requirements. This research project intends to construct a PCNN deep model that would better mimic the brain visual mechanism, by integrating the advantages of the deep Convolution Neural Networks (CNN) in the characteristic expression, and the mechanism of the Pulse Coupled Neural Network (PCNN) in mimicking the brain visual cortex in the synchronization pulses coupled. Furthermore, the proposed PCNN deep model would be applied to the automatic feature extraction of the breast cancer mitosis cells. In addition, this research project will also explore the following issues: the method to segment interest mitotic nucleus regions based on PCNN, the way to train the convolutional parameters of the proposed PCNN deep model, the way to adaptively set the values of other PCNN parameters, as well as the method to extract PCNN deep learning features from the interest regions. The aim is to improve the accuracy rate of mitosis detection in breast histopathological images, and assist pathologists in offering more objective diagnosis information. Therefore, this research project would serve as a theoretical foundation and technical support for constructing the computer-aided diagnosis system.

组织病理图像有丝分裂细胞计数是乳腺癌病理分级的关键因素。人工检测有丝分裂细胞常因工作量大出现误诊漏诊，且不同病理学家之间甚至同一病理学家不同时期的诊断意见也不一致，因此迫切需要借助计算机图像自动分析技术对乳腺组织病理图像进行量化分析和评估。.针对现有自动分析方法准确率较低无法满足临床需求的现状，本项目将有机结合深度卷积神经网络（CNN）的深度特征表达结构和脉冲耦合神经网络（PCNN）的同步脉冲发放与耦合机理，拟构建一种更能模仿大脑视觉信息处理的PCNN深度模型，并将之应用于乳腺癌有丝分裂细胞特征的自动提取中。此外，还将探索基于PCNN的有丝分裂细胞核感兴趣区域分割、PCNN深度模型参数训练与自适应设置、以及感兴趣区域PCNN深度学习特征的提取等方法，以期进一步提高乳腺组织病理图像有丝分裂细胞检测的准确率，为病理学家的诊断提供客观的辅助信息，为计算机辅助诊断系统的构建提供理论基础和技术支持。

医学组织病理图像中普遍存在细胞核粘连现象，这是造成细胞核区域难以准确分割，病变细胞数量难以精准统计的主因。本项目提出了一种精准、高效检测组织病理图像粘连细胞核边缘的Caps-Unet模型，该模型提出用四个不同尺度的卷积单元和一个连接单元所组成的胶囊层（Caps-Layer）替换原U-Net模型的卷积层。实验表明，Caps-Unet模型能检测出U-Net模型检测不到的粘连细胞核边缘；相比于U-Net模型，损失率下降了约5%，评价指标Dice系数提高了约2%，且收敛速度更快。.为构建拟融合深度CNN特性的深度脉冲耦合神经网络（深度PCNN）模型，需要深入研究CNN特性。本项目以LeNet-5为基础，分别以Google街景门牌号数据集SVHN、美国手势数据集ASL、德国交通标志数据集GTSRB、人脸数据集MTFL和GENKI-4K为研究对象，探索适于各数据集的最优CNN模型和算法。重要结果和数据为：(1)提出KM-Net模型，利用K-Means算法替代随机数法初始化CNN的首个卷积权值矩阵。在SVHN上仅需训练0.95h达到93.80%正确率；在ASL上只需迭代30次可达99.55%准确率。(2)进一步融合CNN跨层特征，提出WCF-CNN模型，SVHN上仅需训练2.2h识别率即升至95.6%。(3) 提出基于改进LeNet-5模型和SVM分类器的微笑识别算法，在MTFL和GENKI-4K上准确率分别为87.3% 和86.3%。(4)提出先粗分类后细分类的二级改进LeNet-5的交通标志识别算法，在GTSRB上正确率可达91.76%。.针对传统深度CNN模型不考虑生物脉冲时间编码特性，需大量有监督训练数据才能保证其性能的特点，本项目提出了一种基于SPCNN时间编码的无监督P-Spiking深度神经网络（P-SDNN）模型，其权重参数依据STDP规则更新。P-SDNN模型在脉冲深度神经网络（SDNN）基础上引入一个SPCNN时间编码（Tem-coding）层，用于对经过DOG处理后的图像进行时间编码。该层可对不同输入图像，自适应地生成不同数量的时间脉冲图，而非指定相同数量，从而合理分配计算资源。实验表明， P-SDNN模型在公认数据集Caltech101和MNIST数据集上的图像分类性能优于无SPCNN时间编码的SDNN模型；且相比于传统深度模型，所需训练数据大为减少。