Topology-Aware GAN (TopoGAN): Transforming Medical Imaging Advances

EasyChair Preprint no. 11474

Abstract

Generative Adversarial Networks (GANs) have gained prominence in medical imaging due to their ability to generate realistic images. Traditional GANs, however, often fail to capture intricate topological features such as holes and connectivity components in real images. This study applies TopoGAN, a recently-developed model tailored for medical imaging. TopoGAN dynamically learns and incorporates topological features like connectedness and loops, addressing a real-world medical data augmentation problem. Utilizing a topological GAN loss function based on Persistent Homology and a new success metric, TopoGAN minimizes the topological discrepancy between synthetic and actual images. Experimental results, showcasing a Wasserstein distance of 0.0021 and a Dice coefficient of 0.995, highlight the model’s efficacy in producing qualitatively rich synthetic images. This approach not only improves the realism of generated images but also enhances performance in downstream tasks such as image segmentation, offering a groundbreaking solution with significant implications for medical image analysis, diagnosis, and treatment planning.

Keyphrases: Generative Adversarial Network, mathematics, Medical Imaging, persistent homology, topological data analysis

@Booklet{EasyChair:11474,
  author = {Hrishi Patel and Colleen Farrelly and Quincy Hathaway and Jennifer Z Rozenblit and Deepa Deepa and Yashbir Singh and Ashok Chaudhary and Yassine Himeur and Wathiq Mansoor and Shadi Atalls},
  title = {Topology-Aware GAN (TopoGAN): Transforming Medical Imaging Advances},
  howpublished = {EasyChair Preprint no. 11474},

  year = {EasyChair, 2023}}