精准医疗带来了更加清晰的医学图像的需求，但长期的低精度医学成像设备现在仍在一二级医疗机构中广泛应用，其生成的图像具有低精度高噪声的特点，故此应用图像处理的方式实现核磁图像（全称，简称）有效降噪具有非常重要的研究意义。

作为最新的生成模型方法，扩散模型在底层的图像翻译应用中取得了显著成果，实现了从高层次的细节到生成示例的多样性。本文研究了使用改进后的去噪扩散概率模型对脑MRI三维降噪方法，面对高精度脑MRI图像实现应用，相关的主要创新性工作为：

1、在研究中设计去噪扩散概率模型，可实现与其他基于似然的模型竞争的对数似然，即使是在像ImageNet这样的高多样性数据集上，方法也依然有效。为了更紧密地优化变分下界（variational lower-bound VLB），使用简单的重新参数化和混合学习目标来学习反向过程方差。

2、面对MRI降噪工作，我们提出了一个混合目标函数。基于此目标函数，可获得了比直接优化对数似然获得的更好的对数似然，并发现后一个目标在训练过程中有更多的梯度噪声。通过研究希望证明应用简单的重要采样技术可以减少梯度噪声，并允许我们实现比使用混合目标更好的对数似然。

3、构建残差模块，在将学习到的方差纳入设计的模型后，以更少的步骤从实现模型中取样，而样本质量的变化很小。相关研究可以有效地解决传统DDPM需要数百个正向通道来产生良好的样本，使得正向通道采样数量降低，也可以获得良好的样本，从而加快了在实际应用中使用的采样速度。

面向实际应用，构建了脑MRI降噪图像平台。通过与GAN模型、传统去噪扩散模型、transform与VAE的对比实验表明，去噪图像重构精度，模型的运行速度等方面都有显著提高，同时算法具有良好的鲁棒性，可以基本满足脑MRI三维降噪的需要。

Precision medicine has brought the demand for clearer medical images, but long-term low-precision medical imaging equipment is still widely used in first-and second-class medical institutions, and the images generated by it have the characteristics of low precision and high noise. Therefore, it is of great research significance to apply image processing to realize effective noise reduction of nuclear magnetic resonance images.

As the latest generation model method, diffusion model has made remarkable achievements in the application of image translation at the bottom, realizing the diversity from high-level details to generation examples. In this paper, the improved denoising diffusion probability model is used to denoise brain MRI in three dimensions. Facing the application of high-precision brain MRI images, the main innovative works are as follows:

1. In the research, the denoising diffusion probability model can be designed, which can compete with other likelihood-based models. Even on high diversity data sets like ImageNet, the method is still effective. In order to optimize the variational lower-bound VLB more closely, simple reparameterization and mixed learning objectives are used to learn the inverse process variance.

2. Facing the noise reduction of MRI, we put forward a mixed objective function. Based on this objective function, a better log likelihood can be obtained than that obtained by directly optimizing log likelihood, and it is found that the latter target has more gradient noise in the training process. Through research, it is hoped to prove that the application of simple important sampling technology can reduce gradient noise and allow us to achieve better log-likelihood than using mixed targets.

3. Build a residual module, and after incorporating the learned variance into the designed model, sample from the realized model in fewer steps, with little change in sample quality. Related research can effectively solve the problem that traditional DDPM needs hundreds of forward channels to produce good samples, which reduces the number of forward channel samples and can also obtain good samples, thus speeding up the sampling speed used in practical applications.

For practical application, a brain MRI denoising image platform is constructed. The comparative experiments with GAN model, traditional denoising diffusion model, transform and VAE show that the accuracy of denoising image reconstruction and the running speed of the model have been significantly improved, and the algorithm has good robustness, which can basically meet the needs of three-dimensional noise reduction of brain MRI.