| Abstract: |
High-precision medical image processing tasks, such as tumor detection and segmentation, generally require large amounts of training data. However, constructing large-scale medical image datasets is challenging due to privacy and ethical constraints, as well as the rarity of certain diseases. To address this issue, data augmentation using generative models, including Generative Adversarial Networks (GANs), has been extensively investigated, and their effectiveness has been demonstrated in various medical imaging tasks (Frid-Adar et al., 2018). Nevertheless, GAN-based approaches often suffer from training instability, making it difficult to consistently ensure the quality and diversity of generated images. In recent years, advances in diffusion models have led to an increasing number of studies exploring diffusion-based data augmentation for medical image analysis, and more stable generation frameworks are gradually being established (Nazir et al., 2025). However, in the context of medical image segmentation, several practical aspects remain insufficiently organized, including (i) operation on small-scale datasets, (ii) reproducible procedures for prompt design, and (iii) implementations that explicitly preserve anatomical structures. In this study, we focus on these practical limitations and evaluate diffusion-based data augmentation, specifically using Stable Diffusion, for medical image segmentation. We present a reproducible augmentation workflow encompassing prompt generation, sample selection, anatomy-preserving image generation, and quantitative evaluation. Medical image datasets are augmented using Stable Diffusion, and multiple segmentation models, such as U-Net, are trained on the augmented datasets. The effectiveness of the proposed workflow is quantitatively evaluated using the Intersection over Union (IoU) and Dice coefficient across multiple datasets. |