基于引入注意力机制扩散模型的儿童头颅CT生成研究OA

In view of the difficulty in obtaining medical images and the deficiencies of traditional generative models,an improved diffusion model CBOM-MDDPM(Diffusion Probabilistic Model Enhanced with Convolutional Block Attention Module and MobileNetV3)based on depth-separable convolutional MobileNetV3 and the attention mechanism is proposed.The 463 cases of children's skull CT(Computed Tomography)data provided by Xinhua Hos-pital Affiliated to Shanghai Jiao Tong University are taken as the image training set.By integrating the bneck module in MobileNetV3 into the UNet(U-shaped Network)encoding part of the diffusion model and introducing the CBAM attention mechanism and other strategies in its downsampling part,the computational burden is reduced,and the im-age generation quality and sampling speed are improved.The experimental results show that the FID(Frechet Incep-tion Distance)and IS(Inception Score)indicators of the proposed model are 38.92±1.06 and 2.25±0.035 respec-tively,and the quality of the generated images is superior to that of the traditional GAN(Generative Adversarial Net-work)model.Compared with the DDPM(Denoising Diffusion Probabilistic Mode)model,the number of parameters of the CBAM-MDDPM model has decreased by 38 percentage points,and the processing speed has increased by 58 percentage points,which proves the effectiveness of the proposed algorithm.