[2M4-CC] Late Breaking Research Session - B

[2M4-CC-04] Generative Adversarial Network driven Synthetic Augmentation of Confocal Immunofluorescence Image Segmentation

Daiki Katsuma¹, Hirhoaru Kawanaka¹, V.B. Surya Prasath², Bruce J. Aronow² (1. Mie University, 2. Cincinnati Children’s Hospital Medical Center)

Human lung development requires complex gene and cell interactions, and lung development can be studied at gene, cellular, and molecular levels. Availability of multimodal imaging data of the lung can help us visualize protein and cell localization in various lung structures. In particular, confocal Immunofluorescence (IF) images can be used in lung development modeling, and the authors require multi-class segmentation as a necessary requirement. It is, however, difficult to get a sizable number of training images in the multi-class segmentation with recent deep learning models, e.g. the convolutional neural network (CNN) methods. Here, in order to improve the overall accuracy of multi-class segmentation in lung confocal IF images, we carried out the data augmentation with Generative Adversarial Network (GAN). By creating high fidelity synthetic images generated by GAN, it is expected that the synthetically generated images can be mixed with the original images to increase the accuracy of the automatic segmentation.
In this study, we investigate the effects of changing generation models and the number of synthetic images in the training segmentation models. Our experimental results indicate that we can generate better quality synthetic image sets by using GauGAN. Also, the more significant number of synthetically generated images can be of use in deep CNN based segmentation models. At the moment, our approach were up 13.1％ on average in 6 classes classification compared to the case without adding the synthetic images. In the future, we should implement and train other segmentation models with synthetic images. We hope to be satisfied with sufficient accuracy for further lung analysis.