Super-Resolution Reconstruction of Medical Images Based on Improved Generative Adversarial Networks

In the process of acquiring medical images, the quality of the acquired medical images may be poor due to various reasons, thus leading to the doctor's wrong judgement of the patient's condition. In order to solve the problem of unclear texture and blurred edges in medical image storage, an improved generative adversarial network for super-resolution reconstruction of medical images is proposed. Inspired by the Res2Net network, the generator of SRGAN is improved and a multi-scale feature fusion module is designed as the backbone feature extraction network of the generator. This module represents the image multiscale features at the granularity level and expands the receptive field of each network layer to fully extract the feature information of the image. Meanwhile, the EMA attention is embedded into the multiscale feature fusion module, which improves the ability of the model to process features and enhances the feature representation by reorganising the dimensionality of the channels and the batch dimensions, and using the interaction of the different dimensions to capture the connection between image pixels. The results show that at a magnification of 4, this paper's method has better results in objective evaluation criteria compared to SRResNet and SRGAN. It can be seen that the algorithm in this paper can enhance the texture details of medical images and obtain better visual effects.