Medical image local augmentation via text- and mask-guided diffusion model.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computer Vision

Background:

• Medical image scarcity hinders intelligent analysis and precision medicine.
• Current data augmentation methods lack control over local image details.
• Advanced techniques are needed to overcome data limitations in medical AI.

Purpose of the Study:

• To propose a text- and mask-guided local augmentation method for medical images.
• To enhance the diversity and quality of synthesized medical images.
• To enable fine-grained control over local image regions for data augmentation.

Main Methods:

• Utilized a pre-trained MedSAM model for precise segmentation of target regions, generating masks.
• Developed semantically relevant and task-specific text prompts for diverse medical imaging data.
• Integrated masks and text prompts into a diffusion generative model for controlled local perturbations.

Main Results:

• Achieved significant reduction in local structural similarity (97.9% on chest x-ray, 103.3% on pelvic CT, 42.2% on brain CT) within masked regions.
• Demonstrated effective modulation of structural features in local regions.
• Maintained global texture consistency in generated synthetic medical images.

Conclusions:

• Presents a new pathway for controlled data augmentation in medical imaging.
• Facilitates advancements in intelligent medical image analysis.
• Lays the groundwork for future research in fine-grained medical image generation.