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

Medical imaging is crucial for artificial intelligence-based clinician decision-making. However, learning anatomical features from limited samples poses a challenge. To address this issue, recent studies have employed diffusion models to generate medical imaging data, demonstrating the potential for high-quality medical image generation through deep learning. In this paper, we propose a high-quality cervical MRI image generation method called the Cervical Spine MRI Diffusion Probabilistic Model (CSM-DPM). Considering the complexity of neck MRI image data, our method uses a hybrid of standard Gaussian noise and point noise obtained by sampling within 2D Gaussian noise fields to approximate the true distribution of the image data. Meanwhile, the cosine noise schedule is used to make the morphology of the generated vertebral blocks and other focal areas more visually natural and clear. Furthermore, to enhance the noise prediction capabilities of UNet in DDPM, we devise the Asa-ResUNet module, which incorporates an asymmetric attention mechanism. This mechanism includes spatial attention on different ResUNet layers to improve feature extraction and incorporates high-level semantic information. We further enhance the stability and robustness of the Asa-ResUNet by using an exponential weighting strategy (EMA). Experiments demonstrate that our method produces cervical spine MRI images with FID values that are up to 15.79%, 52.61%, and 46.57% lower than those produced by DDPM, DDIM, and F-PNDM, respectively, indicating superior image quality. Segmentation experiments confirm that the generated images can enhance the overall performance of segmentation models when used for training.