An Example for Domain Adaptation Using CycleGAN

Cycle-Consistent Adversarial Network (CycleGAN) is very promising in domain adaptation. In this report, an example in medical domain will be explained. We present struecture of a CycleGAN model for unpaired image-to-image translation from microscopy to pseudo H&E stained histopathology images.

💡 Research Summary

This paper presents a practical application of Cycle‑Consistent Generative Adversarial Networks (CycleGAN) for virtual staining, translating fluorescence light‑sheet microscopy images into pseudo‑H&E histopathology images without requiring paired training data. The authors first motivate the need for such translation: fluorescence images preserve fine structural detail but are not readily interpretable by pathologists who rely on the colour and texture cues of H&E‑stained slides. Traditional paired image‑to‑image translation methods are infeasible in this context because acquiring perfectly aligned fluorescence‑H&E pairs is technically challenging and time‑consuming.

To address this, the authors adopt a standard CycleGAN architecture consisting of two generators (G_A2B for fluorescence→H&E and G_B2A for the reverse) and two discriminators (D_A for the fluorescence domain, D_B for the H&E domain). Each generator follows a ResNet‑based encoder‑decoder design: an initial 7×7 convolution with reflection padding (64 filters), two stride‑2 down‑sampling layers that increase the channel depth to 256, nine residual blocks equipped with instance normalization and ReLU activations, followed by two transposed‑convolution up‑sampling layers and a final 7×7 tanh‑activated convolution to produce RGB outputs in the range