An end-to-end CNN and LSTM network with 3D anchors for mitotic cell detection in 4D microscopic images and its parallel implementation on multiple GPUs

Titinunt Kitrungrotsakul, Xian-Hua Han, Yutaro Iwamoto, Hideo Yokota, Satoko Takemoto, Sari Ipponjima, Tomomi Nemoto, Wei Xiong, Yen-Wei Chen

doi:10.1007/s00521-019-04374-8

The detection and observation of mitotic event are the key to studying the behavior of the cell and used to examine various diseases. The existing cell detection methods are performed on two-dimensional images with time sequence. However, the complexity of mitotic and normal cells and the orientation of the mitosis generate high false positive when using 2D methods. On the other hand, 3D methods can perform higher performance than 2D methods but also face the problem of overfitting due to the limit of training data. With those problems, we propose a 2.5-dimensional convolutional neural network with convolutional long short-term memory to extract the information time sequence and combined with 3D anchors to gather the spatial information for final mitotic detection. Furthermore, we also propose the method with a parallel model on multi-GPUs to speed up the detection time. Compared with state-of-the-art methods, our method can reach high precision and also recall rate with detection time is speed up about 1.9 times by the use of the parallel model on 4GPUs.

An end-to-end CNN and LSTM network with 3D anchors for mitotic cell detection in 4D microscopic images and its parallel implementation on multiple GPUs

Bibliographic Information

Search this article

Description

Journal

Citations (1)*help

References(24)*help

Related Projects

Details 詳細情報について

Export

Report a problem