ECP-GAN: Generating Endometrial Cancer Pathology Images and Segmentation Labels via Two-Stage Generative Adversarial Networks
Endometrial cancer is one of the most common tumors of the female reproductive system and ranks third in the world list of gynecological malignancies that cause death. However, due to the privacy and complexity of pathology images, it is difficult to obtain pathology images and corresponding annotation, which affect the accuracy of pathology image segmentation and analysis. To address this issue, this paper proposes a two-stage endometrial cancer pathology images- and labels-generating network, which can generate pathology images and corresponding segmentation labels. In the images-to-images network, a pathological style feature information fusion normalization module is proposed, which decouples the original style feature into style feature vectors to provide independent style feature information. In the images-to-labels network, a pathological prior features guidance loss block is proposed, which improves the ability of the model in feature extraction, the segmentation label-generation accuracy, and the boundary sensitivity to the target region. Training ECP-GAN in the solid tumor endometrial cancer pathological dataset, by physician recognition and experiments on the medical image segmentation tasks, shows that the ECP-GAN network generates realistic images and significantly improves the accuracy of segmentation tasks, which improves about 20% of the segmentation evaluation indicators. Through comparative analysis, the experimental results show that the proposed method effectively improves the robustness and accuracy of the model in segmentation tasks. Particularly when dealing with the complex morphological features of pathology images, this method enhances the model's ability to adapt to various changes, significantly improving.