基于条件扩散模型的结直肠图像染色归一化

Diffusion model-based staining normalization for colorectal image

由于现有染色归一化方法无法准确提取结直肠病理图像的复杂结构特征,导致丢失部分结直肠病理图像的结构信息,无法生成高质量的染色归一化结直肠病理图像。为解决该问题,提出一种基于条件扩散模型的结直肠病理图像染色归一化方法,该方法包括条件扩散模型和图像特征重建。在条件扩散模型中,首先,使用马尔科夫链前向过程对结直肠病理原图像进行加噪声。然后,将噪声图像和条件图像输入到增强去噪网络中进行去噪,在这过程中利用增强激活模块,学习结直肠病理图像的深层特征,捕获更多的图像上下文信息。在编码器和解码器之间引入跳跃连接空间注意力模块,准确提取结直肠病理图像的位置空间信息。在图像特征重建中,设计金字塔特征提取模块,提取多尺度条件图像与生成图像的特征,并构建重建损失函数,优化整个网络的性能。实验结果表明,与现有方法相比,所提染色归一化方法在公共数据集GlaS和CRAG上的能生成质量更高的染色归一化结直肠病理图像。

Existing staining normalization methods are unable to accurately extract the complex structural features of colorectal pathological images,resulting in the loss of partial structural information and the inability to generate high-quality staining-normalized colorectal pathological images.To address this issue,a staining normalization method for colorectal pathological images based on a conditional diffusion model is proposed.The proposed method includes conditional diffusion model and image feature reconstruction.In conditional diffusion model,firstly,the Markov chain forward process is employed to add noise to the original colorectal pathological images.Then,the noisy images and conditional images are input into an enhanced denoising network for denoising.During this process,an enhanced activation module is utilized to learn the deep features of the colorectal pathological images and capture more contextual information.A skip-connection spatial attention module is introduced between the encoder and decoder to accurately extract the positional spatial information of the colorectal pathological images.Finally,a pyramid feature extraction module is designed to extract the features of the multi-scale conditional images and generated images,and a reconstruction loss function is constructed to optimize the performance of the entire network.Experimental results demonstrate that compared with existing methods,the proposed staining normalization method can generate higher-quality staining-normalized colorectal pathological images on public datasets GlaS and CRAG.