Year: 2026
Type: article
Abstract: Diffusion models are powerful deep generative models that achieve high-quality image and video generation, and have also been applied to medical image segmentation. However, conventional diffusion models rely on stochastic generation processes that explore the entire data distribution, which may not be well suited to segmentation tasks where a unique correct output is required for each input image. In this study, we focus on Rectified Flow, which reformulates the stochastic diffusion process as a deterministic ordinary differential equation. By learning a one-to-one mapping from noise to output, Rectified Flow preserves the generative capability of diffusion models while providing a deterministic generation process that is consistent with the task structure of medical image segmentation. Based on this idea, we propose a Rectified Flow-based diffusion model for medical image segmentation. Furthermore, we introduce a Transformer-based architecture with cross attention between the conditional medical image and the segmentation image to accurately capture complex spatial relationships. Experiments on a public skin lesion dataset compare the proposed method with existing segmentation approaches in terms of segmentation accuracy, computational cost, and model size, demonstrating its effectiveness.
DOI: 10.5687/iscie.39.279
Institution: Ritsumeikan University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7230894873
