提出一种利用细胞时域演化特征来表示有丝分裂的方法。首先,该方法从细胞有丝分裂序列的每帧图像中提取三种不同特征,即通用搜索树(GIST)、尺度不变特征变换(SIFT)、卷积神经网络(CNN),针对提取的每种特征,用池化方法在空间和时间维度上对其进行处理;然后,将处理后的一系列池化特征组合成一个单一的特征向量来表示该细胞有丝分裂事件的最终特征;最后,将组合后的特征向量作为分类器的输入,利用传统的机器学习方法,即支持向量机(SVM)来处理有丝分裂事件的识别问题。实验结果表明所提方法在精度和召回率方面均优于传统方法,可以更好地应用于细胞有丝分裂检测中。

Herein, a method to represent a mitosis event is proposed by using the feature of cell evolution in the time domain. First, three kinds of features are extracted for each frame of the mitotic sequence, i.e., the generalized search tree, scale invariant feature transformation, and convolutional neural network. Each series of extracted features is handled using the pooling method in spatial and temporal dimensions. Subsequently, the processed series of pooling features are combined into a vector to represent the final mitotic event characteristics. Finally, the combined feature vector is used as the classifier input, and the traditional machine learning method of support vector machine is used to address the mitotic recognition problem. The experimental results denote that the proposed method is superior to the traditional method with respect to the precision and recall rate and is more appropriate for mitosis detection applications.