Abstract
Histopathology image synthesis aims to address the data shortage issue intraining deep learning approaches for accurate cancer detection. However,existing methods struggle to produce realistic images that have accurate nucleiboundaries and less artifacts, which limits the application in downstreamtasks. To address the challenges, we propose a novel approach that enhances thequality of synthetic images by using nuclei topology and contourregularization. The proposed approach uses the skeleton map of nuclei tointegrate nuclei topology and separate touching nuclei. In the loss function,we propose two new contour regularization terms that enhance the contrastbetween contour and non-contour pixels and increase the similarity betweencontour pixels. We evaluate the proposed approach on the two datasets usingimage quality metrics and a downstream task (nuclei segmentation). The proposedapproach outperforms Sharp-GAN in all four image quality metrics on twodatasets. By integrating 6k synthetic images from the proposed approach intotraining, a nuclei segmentation model achieves the state-of-the-artsegmentation performance on TNBC dataset and its detection quality (DQ),segmentation quality (SQ), panoptic quality (PQ), and aggregated Jaccard index(AJI) is 0.855, 0.863, 0.691, and 0.683, respectively.