KUnet: Microscopy image segmentation with deep unet based convolutional networks
Journal
Conference Proceedings - IEEE International Conference on Systems, Man and Cybernetics
Journal Volume
2019-October
Pages
3561-3566
Date Issued
2019
Author(s)
Chang S.W
Abstract
Due to the temporal behavior of living cells, the microscopy sequences analysis is indeed a challenging task. Unet architectures are normally considered as one of the most powerful tool for segmentation of biomedical images. However, there is still no clear way to deal with the temporal and spatial characteristics of the time-series cell images. This study uses the weight pre-trained on a large scale of data as initial weights and Convolutional Long Short Term Memory (CLSTM), which replace most of the pure convolution operation in the original Unet, to obtain better performance than many significant network. We compared eight UNet encoder network architectures: VGG11, VGG13, VGG16, VGG19, Resnet18, and Densenet121, Inceptionv3 and Incetionresnetv2. Among all the networks, it is the architecture of VGG13 that obtain the best segmentation results on testing data (a public available data set). We then evaluated the performance of these models with multiple training, validation and testing. Two evaluation methods (intersection over union, IoU and Error metrics) are adopted to solidify the work of the improvement. In this paper, we firstly discovered the most powerful structure for encoder of Unet through plentiful experiments and comparison of multiple deep learning models. Secondly, we further successfully enable the best model to perform spatiotemporal encoding. For evaluation, the comparison with other significant Unet-based and FCN-based network are made eventually. ? 2019 IEEE.
Subjects
Convolution; Deep learning; Network architecture; Scales (weighing instruments); Signal encoding; Statistical tests; Biomedical images; Convolutional networks; Evaluation methods; Learning models; Segmentation results; Sequences analysis; Temporal and spatial; Temporal behavior; Image segmentation
SDGs
Type
conference paper