Hybrid-noise generative diffusion probabilistic model for cervical spine MRI image generation.

Area of Science:

• Medical imaging
• Artificial intelligence
• Deep learning

Background:

• Medical imaging is vital for AI-assisted clinical decision-making.
• Generating high-quality medical images from limited data is challenging.
• Diffusion models show promise for realistic medical image synthesis.

Purpose of the Study:

• To propose a novel method for generating high-quality cervical spine MRI images.
• To address the challenge of learning anatomical features from limited medical imaging data.
• To improve AI-based clinician decision-making through enhanced image generation.

Main Methods:

• Developed the Cervical Spine MRI Diffusion Probabilistic Model (CSM-DPM).
• Employed a hybrid noise approach (standard Gaussian and point noise) for accurate data distribution approximation.
• Utilized a cosine noise schedule for natural and clear generation of anatomical structures.
• Introduced the Asa-ResUNet module with an asymmetric attention mechanism to improve noise prediction.
• Applied an exponential moving average (EMA) strategy for enhanced model stability and robustness.

Main Results:

• CSM-DPM achieved superior image quality, demonstrated by significantly lower FID scores compared to DDPM, DDIM, and F-PNDM.
• Generated cervical spine MRI images were visually natural and clear, particularly in anatomical details.
• Segmentation experiments showed that the synthetic images improve the performance of segmentation models.

Conclusions:

• The proposed CSM-DPM effectively generates high-quality cervical spine MRI images.
• The method enhances the learning of anatomical features for AI applications in medicine.
• This work contributes to advancing deep learning-based medical image synthesis for clinical decision support.