Dilated-residual U-Net for Optical Coherence Tomography noise reduction and resolution improvement

Optical coherence tomography (OCT) is a non-invasive 3D imaging technique that provides high-resolution images, and has been extensively used in biomedical research and clinical studies. Although micrometer resolution is already considered high for biological tissue imaging, the need for even higher resolution remains constant. Improving the resolution of OCT images can reveal previously unseen microstructures, which can aid in achieving more accurate diagnoses. Currently, the resolution of OCT images is primarily constrained by speckle noise and spectral bandwidth limitations. We have achieved simultaneous suppression of speckle noise and resolution improvement in OCT images in our previous work. However, traditional methods based on prior optimization iteration have a high computational cost, which limits its applicability. In this paper, we propose an improved deep learning model called DRUNET (Dilated-Residual U-Net) to achieve noise reduction and resolution improvement simultaneously. The model incorporates dilated convolution and residual learning to enhance the learning capacity of the U-Net. In addition, we apply a simple yet effective attention module called Convolutional Block Attention Module (CBAM) to improve DRUNET performance. We evaluate the performance of the DRUNET model in denoising and improving resolution on two types of OCT images. The experimental results demonstrate the effectiveness of the proposed model, which enables us to batch process poor-quality OCT images quickly without requiring any parameter fine-tuning under time constraints.