级联稀疏卷积与决策树集成的病理图像细胞核分割方法

数字病理图像分析对于乳腺癌、肾癌等良恶性分级诊断具有重要意义, 其中细胞核的形态测量是病理量化分析的关键. 然而, 由于病理图像背景复杂, 细胞核高密度分布、细胞粘连等, 个体细胞核精准分割是一个挑战性问题. 本文提出一个级联稀疏卷积与决策树集成学习的细胞核分割模型. 该模型由稀疏可分离卷积模块和集成决策树学习的正则化回归模块堆叠级联组成, 其中: 前者采取秩-1张量分解学习机制, 可分层抽取细胞核的多尺度方向分布式抽象特征; 而后者采取随机采样、树剪枝以及正则化回归机制提升逐像素回归分类能力. 相比于现有深度学习模型, 该模型无需非线性激活和后向传播计算, 参数规模较小, 可实现端到端的学习. 通过乳腺、前列腺、肾脏、胃和膀胱等多组病理图像的分割实验表明: 该模型能够实现复杂数字病理图像中的高密度细胞核的快速个体目标检测和分割, 在Jaccard相似性系数、F1分数和平均边缘距离三个指标上均优于目前CNN2、CNN3和U-Net等深度学习方法, 具有较好应用前景.

Cascade Sparse Convolution and Decision Tree Ensemble Model for Nuclear Segmentation in Pathology Images

The quantitative analysis of digital pathology images plays a significant role in the diagnosis of benign and malignant diseases such as breast and prostate cancer,in which nuclear morphology measurement serve as a basis of quantitative analyses.However,due to the complex background of pathology images,dense distributions of nuclei,and nucleus adhesions,accurate segmentation of individual nuclei remains a challenging problem.In this paper,we propose a new method to automatically segment nuclei from digital pathology images with cascade sparse convolution and decision tree ensemble(CscDTE) model.In particular,the sparse separable convolution learning module and the decision tree ensemble learning module are stacked in a cascaded manner to form the CscDTE model.The former adopts rank-one tensor decomposition learning mechanism that can extract multiscale and multi-directional distributed abstract features;while the latter employs random sampling,pruning,and regularized regression mechanism to boost per-pixel regression and/or classification performance.Compared with the popular deep neural networks,the proposed CscDTE model does not require nonlinear activation and backpropagation computation,and depends on fewer parameters.Our CscDTE model is trained in a layer-wise manner that can achieve an end-to-end pixelwise learning and fast nuclear detection and segmentation in high-throughput imagery.We demonstrated the superiority of our method in terms of Jaccard index,F1 score,and average boundary distance by evaluating it on the multi-disease state and multi-organ dataset where consistently higher performance was obtained as compared to convolutional neural networks and fully convolutional networks.