Fusing residual dense and attention in generative adversarial networks for super-resolution of medical images

The resolution of clinical medical images may be reduced due to differences in the radiation dose, sampling equipment, and storage methods. Low-resolution (LR) medical images blur lesion feature information and affect the diagnostic accuracy of clinicians. To address this issue, we proposed the fusing residual dense and attention in generative adversarial network (FRDAGAN) for the super-resolution (SR) of medical images.

The FRDAGAN method is based on the generative adversarial network (GAN). First, the feature information of different layers is fully utilized by residual dense blocks to prevent gradient decay. Second, the attention gate (AG) network is used to suppress the noise information and improve the signal-to-noise ratio. Finally, a hybrid loss function is used to prevent vanishing or exploding gradients in the network.

On the Luna16 dataset, using single images, the quantitative SR results, peak signal-to-noise ratio (PSNR), and mean structural similarity index measure (SSIM) were 31.257/0.964, 33.558/0.968, and 34.201/0.882, respectively. While on the brain magnetic resonance imaging (MRI) dataset, those values were 34.220/0.874, 35.735/0.885, and 35.854/0.908, respectively. The proposed method showed obvious enhancement compared to the other methods on the Luna16 and brain MRI test data sets, and the PSNR and SSIM values reached 33.005±0.157, 0.938±0.028, and 35.270±0.183, and 0.889±0.024, respectively. The uniform resource locator (URL) for Luna16 is https://luna16.grand-challenge.org/Download/, and the URL for brain MRI is https://www.kaggle.com/datasets/mateuszbuda/lgg-mri-segmentation.

The FRDAGAN method had better results in terms of the PSNR and SSIM than the other traditional methods, and was more stable in terms of the faster convergence of the loss function. The results showed that the FRDAGAN method is effective and advanced.