Super-resolution reconstruction of OCT images based on frequency and spatial information in adversarial neural networks

<i>Objective.</i>Optical coherence tomography (OCT) has a wide range of applications in the diagnosis and treatment of diseases such as heart and ophthalmic diseases. However, the inherent limitations of imaging hardware, low spatial sampling rates, and noise severely degrade image resolution and fine detail visibility.<i>Approach.</i>We proposed an adversarial neural network that integrates spatial and frequency information for OCT image super-resolution (SR) reconstruction. Two key modules, frequency-convolution-batch normalization-rectified linear unit (FCBR) and PixelShuffler- FCBR, were built to perform frequency and spatial feature extraction and fusion on low resolution images and SR images, respectively.<i>Main Results.</i>Extensive experiments on both coronary artery and ophthalmic OCT datasets demonstrate our model achieves substantial improvements in image detail restoration, frequency domain fidelity, and overall visual quality compared to existing state-of-the-art methods.<i>Significance.</i>In conclusion, by jointly leveraging global frequency components and local spatial features, the proposed method significantly improves texture recovery and structural consistency in OCT images, and may enhance the clinical quantitative evaluation of diseases limited by the resolution of OCT imaging equipment.