Mitosis Detection Based On Structured Feature Representation Model

With the development of artificial intelligence,machine learning has gradually penetrated into the field of biomedical research.However all the biological phenomena were expressed in terms of cell as the basic unit,so that all the exploration on the genetic and physiological function of the organism is inseparable from the in-depth study of cells.Thus it?s necessary to detect and locate mitosis automatically and accurately.In this paper,we proposed methods based on structured time series and applied to feature representation and model learning,this paper mainly introduced the following work:(1)We proposed Pooled Time Series representation,which can track the detail changes of cell state over time,and fuse temporal information and visual feature as the final representation of the whole cell sequence.(2)We proposed 3D Convolutional Neural Network representation,which stacks several consecutive frames to be a cube,and then apply the 3D convolutional kernel to the cube,thus each convolution feature mapping layer and the last layer of multiple adjacent consecutive frames is linked together,which make it captures the cell state transition information between frames well.(3)We introduced Hierarchical Summary Random Field model which is based on structured model learning,which build up a hierarchy dynamically and recursively by alternating sequence learning and sequence summarization.For sequence learning we use CRFs with latent variables to learn hidden spatiotemporal dynamics.This paper expounded the above three kinds of algorithms in detail,and the experimental results show that structured temporal information is an important feature of cell candidate sequence,which can contribute to improve the detection accuracy effectively.