From Normal to Abnormal: Transforming Medical Images with Diffusion Models for Dataset Balancing

Martin Paulikat, Lennart Nauschütte, Martin Simon Kalteis, Witali Aswolinskiy, Achim Schneider, Hermann Bussmann, Christian Aichmüller, Magnus von Knebel Doeberitz

Keywords: diffusion model, digital colposcopy, balancing, cervical screening

Abstract: Digital colposcopy relies on the accurate identification of high-grade lesions in cervical images. This study explores the use of diffusion models to address class imbalance, a common challenge in medical datasets. We propose a method that synthetically generates high-grade lesion features within normal cervical images. This method was evaluated on datasets from Berlin and Cambodia, the latter having a significant scarcity of high-grade lesions. Our approach successfully balanced the dataset and improved diagnostic performance by 5\%.

Submission Number: 89