A Diffusion Probabilistic Model Based on Multi-Residualattention For Medical Image Segmentation
Peng Yang, Chaorong Li, Ling Xu-dong, Fengqing Qin, Zheng Yong, Qiu Lihua · 2023
Diffusion models are currently among the most attention- grabbing generative models, achieving revolutionary applications across multiple domains and demonstrating exceptional performance. In the application of conditional diffusion models to alter specific regions of the output image to match desired features, most methods employ V AE encoders for feature extraction from conditional images. However, V AE encoders face challenges of difficult training and large parameter sizes; simultaneously, the denoising UNet in diffusion models often leads to blurred edges in sampled images. To effectively address these issues, We have introduced ResAtMedDiff, an innovative diffusion model specifically designed for the field of medical image processing, offering a novel approach to enhance the accuracy and efficiency of image analysis. The ResAtMedDiff model consists of two core training networks: the MEUNet conditional encoder and the RsAtUNet encoder-decoder. MEUNet is responsible for efficiently encoding the conditional images and then fusing these encodings with noise image features across multiple scales. This fusion strategy significantly enhances the ability of conditional features to guide the generation of noise images. The MEUN et encoder blends the features of the original conditional image with the Gaussian noise image features of the corresponding layers of the RsAtUN et encoder, followed by the RsAtUNet decoder decoding these fused features. This process achieves the generation of image segmentation masks guided by the features of the original image as medical prior conditions, As a result, this significantly enhances the precision of lesion area segmentation in medical imaging. The experimental findings reveal that the ResAtMedDiff model surpasses existing state-of-the-art segmentation techniques in its performance, particularly in the segmentation of thyroid nodules in ultrasound images and brain tumors in MRI scans, demonstrating its superior efficacy and precision.