Controllable Mask Diffusion Model for medical annotation synthesis with semantic information extraction.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Medical Image Analysis
• Computer-Aided Diagnosis

Background:

• Medical segmentation is vital for AI-driven diagnosis but limited by data privacy regulations.
• High-quality training data, including paired medical and mask images, is essential for segmentation tasks.
• Data augmentation is crucial to overcome data scarcity in medical AI.

Purpose of the Study:

• To propose a novel Controllable Mask Diffusion Model for generating new medical image masks.
• To enable input-driven, controllable mask generation based on semantic information (size, location, count).
• To demonstrate the model's effectiveness in large-scale data synthesis and segmentation task improvement.

Main Methods:

• Developed a Controllable Mask Diffusion Model leveraging mask binary structure for semantic feature extraction.
• Utilized a regressor to apply extracted semantic information as multi-conditional input to a diffusion model.
• Incorporated a technique for analyzing semantic information correlation for large-scale data synthesis.

Main Results:

• Confirmed the model's ability to control and generate new masks based on unseen semantic information.
• Demonstrated improved segmentation task performance when using data augmented with generated masks.
• Achieved superior results on both single-label and multi-label mask augmentation experiments.

Conclusions:

• The Controllable Mask Diffusion Model effectively generates realistic and controllable masks for medical image segmentation.
• The proposed method addresses data scarcity and privacy issues, enhancing segmentation performance.
• This approach shows significant potential for diverse applications across the medical field.