Abstract:
Abstract： Breast cancer is a dangerous malignant tumor. In medicine， human epidermal growth factor receptor 2（HER2）levels are needed to determine the aggressiveness of breast cancer in order to develop a treatment plan， this requires immunohistochemical（IHC）staining of the tissue sections. In order to solve the problem that IHC staining is expensive and time-consuming， firstly， a HER2 prediction network based on mixed attention residual module is proposed， and a CBAM module is added to the residual module， so that the network can focus on learning at the spatial and channel levels. The prediction network could directly predict HER2 level from HE stained sections， and the prediction accuracy reached more than 97.5%， which increased by more than 2.5 percentage points compared with other networks. Subsequently， a multi-scale generative adversarial network is proposed， which uses ResNet-9blocks with mixed attention residuals module as generator and PatchGan as discriminator and self-defines multi-scale loss function. This network can directly generate simulated IHC slices from HE stained slices. At low HER2 level， SSIM and PSNR between the generated image and the real image are 0.498 and 24.49 dB.

Key words: Key words： generative adversarial network, image processing, mixed attention mechanism, category prediction