A data augmentation method based on cycle-consistent adversarial networks for fluorescence encoded microsphere image analysis

摘要

The training process of supervised learning requires a large amount of training data to achieve satisfactory performance. However, the acquisition and annotation from biological images analysis is costly. This article presents a data augmentation method based on Cycle-Consistent Adversarial Networks (CycleGAN) and applied to the annotated image example generation of Fluorescence Encoded Microsphere (FEM) image analysis. A large number of synthetic FEM images and corresponding annotations are generated by computer scripts. A forward generator from CycleGAN is trained to transform the synthetic images into the real images domain for the training data augmentation of the Mask Region Convolutional Neural Network (Mask R-CNN). The training results evaluated for different sizes of real/synthetic/transformed FEM image training sets on Mask R-CNN demonstrate the effectiveness of this method. The average precision at the interval over union of 0.50 (AP(0.)(50)) converges to 95.6% and the AP(0.75) reaches 91.8%, and both are about 10% higher than that of the synthetic image set. The experimental results demonstrate the effectiveness of this method in annotated FEM image augmentation. (C) 2019 Published by Elsevier B.V.