Optical coherence tomography (OCT) imaging shows great potential in clinical practice because of its noninvasive nature. However, two critical issues affect the diagnostic capability of OCT imaging. The first problem is that the interferential nature of OCT imaging produces interference noise, which reduces contrast and obfuscates fine structural features. The second problem is caused by the low spatial sampling rate of OCT. In fact, in clinical diagnosis, the use of a lower spatial sampling rate is a method to achieve a wide field of vision and reduce the impact of unconscious movement. Therefore, most OCT images obtained in reality are not optimal in terms of signal-to-noise ratio and spatial sampling rate. There are significant differences in the texture and brightness of the retinal layer in patients, as well as in the shape and size of the lesion area, so traditional models may not be able to reliably reconstruct the pathological structure. To obtain high peak signal-to-noise ratio (PSNR) and high-resolution B-scan OCT images, it is necessary to develop sufficient methods for super-resolution reconstruction of OCT images. In this paper, an improved OCT super-resolution image reconstruction network structure (PPECA-SRGAN) was proposed.
In this paper, a PPECA-SRGAN network based on generative adversarial network (GAN) was proposed. The network model includes a generator and a discriminator. A PA module was added between the residual blocks of the generator to increase the feature extraction capability of OCT retinal image reconstruction. In addition, a PECA module was added to the discriminator, which is an improvement of the pyramid split attention network (PSANet) and can fully capture the spatial information of multi-scale feature maps. First, we used two data sets to test a training set of 1000 images and a test set of 50 images, respectively. The data set was imported into the preprocessing module, and the low-resolution image was obtained through four down-sampling processes. Then, the generator was used to train the model to generate high-resolution images from low-resolution images. When the discriminator could not distinguish the authenticity of the images, it indicated that the generation network generated high-resolution images. Finally, the image quality was evaluated using the structural similarity index measure (SSIM) and PSNR.
The super-resolution index evaluation results of PPECA-SRGAN and the other three models were compared, as well as the final reconstruction effect images. In general, PPECA-SRGAN’s reconstruction effect was better than SRResNet; however, for the restoration of the image details, the image quality of the PPECA-SRGAN network reconstruction was more in line with the satisfaction degree of human vision. Compared with SRResNet, SRGAN, and ESRGAN, the SSIM indexes of PPECA-SRGAN were 0.090, 0.028, and 0.016 higher and the PSNR indexes were 2.15 dB, 0.71 dB, and 0.47 dB higher, respectively. The good reconstruction effect of PPECA-SRGAN was due to the addition of the attention mechanism called path aggregation network (PANet) and the proposed attention mechanism named PECA, both enhancing the capture of OCT retinal image features and the reconstruction of details. The PECA module was composed of pyramid splitting and extracting features, with the use of ECANet to fuse multi-scale information. PANet can effectively reduce image noise, such as compression artifacts. This makes our model better than the SRGAN network and other traditional networks. Therefore, the application of the proposed model in OCT image super-resolution reconstruction has been verified, and its performance has been improved compared with other reconstruction algorithms.
The PPECA-SRGAN network structure proposed in this paper is an improved model of the SRGAN network for super-resolution reconstruction of retinal OCT B-scan images. We conducted training and verification on MICCAI RETOUCH data set and data collected by Wenzhou Medical University to solve the problems of low-resolution and few details of images collected by OCT. We used advanced GAN to improve the super-resolution reconstruction of OCT images, and the SRGAN network was improved due to the difference in reconstruction between medical images and natural images. Firstly, a PANet module was added between the residual blocks of the generator to extract multi-scale feature relations by pyramid structure and suppress unnecessary artifacts. Then, the PECA module was inserted into the discriminator to effectively combine spatial and channel attentions to learn more image details for the discriminator and obtain richer image pair feature information. The experimental results show that this model is effective and stable in improving the resolution of medical images. Compared with SRResNet, SRGAN, and ESRGAN, the PSNR and SSIM indexes of the reconstructed images were improved by about 3.5% and 5.6%, respectively. In clinical diagnosis, the proposed algorithm can overcome the inherent limitations of low-resolution imaging systems and reconstruct various details lost in the process of image acquisition; the algorithm is easy to integrate and implement. In the future, if higher-quality data sets and lighter algorithms can be obtained, it is possible to further improve the quality of super-resolution reconstruction medical images and make them more applicable in clinical practice.
