Research on high-resolution medical image generation model based on residual convolutional feedforward network
10.19745/j.1003-8868.2025154
- VernacularTitle:基于残差卷积前馈网络的高分辨率医学图像生成模型研究
- Author:
Guang-fa TANG
1
;
Zi-chen LU
;
Xiang-jun MENG
;
Zhuo-kang CHEN
Author Information
1. 广州新华学院信息与智能工程学院,广东 东莞 523000;东莞中科云计算研究院,广东 东莞 523000
- Publication Type:Journal Article
- Keywords:
residual convolutional feedforward network;
high-resolution medical image;
vector quantized generative adversarial network;
deep learning;
image generation
- From:
Chinese Medical Equipment Journal
2025;46(9):1-8
- CountryChina
- Language:Chinese
-
Abstract:
Objective To propose a residual convolutional feedforward network-based high-resolution medical image generation model to enhance the quality for generating high-resolution medical images.Methods Firstly,a vector quantized generative adversarial network(VQGAN)was selected as the benchmark model;secondly,a residual convolutional feedfor-ward network composed of a residual module and a multi-scale convolutional feedforward module was integrated into the the encoder and decoder of VQGAN to form a new architecture;finally,the model propsed was compared with denoising diffusion model(DDM),StyleSwin model,VQGAN moel and SinGAN model to verify its performance,and the ablation experiment was carried out.Results The quantitative evaluation comparison showed that the model proposed behaved well generally with the lowest frechet inception distance(FID)(145.64),the lowest learned perceptual image patch similarity(LPIPS)(0.46)and the peak signal-to-noise ratio(PSNR)(62.63)only lower than that of DDM.Visualized comparison indicated the model proposed had the image generated gaining advantages over the other models in the texture,edges and sharpness of kidney stones and similarity to the real lesion images.The ablation experiment proved that the model proposed behaved better than the VQGAN benchmark model in convergence rate and training stability.Conclusion The proposed model has significant advantages in learning multi-scale texture features and retaining the information on low-dimensional anatomical structures,which can be used for generating high-quality high-resolution medical images.[Chinese Medical Equipment Journal,2025,46(9):1-8]
