A style-aware network based on multi-task learning for multi-domain image normalization

Cervical cell image styles may vary due to factors such as specimen preparation methods and staining schemes. These variations can cause inconsistencies among pathologists and degrade the model performance. Existing staining standardization networks often fail to achieve structure preservation and style approximation. We propose a style-aware network (SA-Net) for multi-domain image normalization to address this issue. SA-Net incorporates a style perception task into the CycleGAN generator to identify different image styles, thus avoiding the need for multiple generators in real-world applications. additionally, We also employ pixel-wise convolutional kernels in the generator to learn only the image style and preserve the image structure. Our experiments demonstrate that SA-Net can effectively enhance the model’s generalization ability and outperform the state-of-the-art methods in multi-style standardization.