Improve deep learning-based reconstruction of optical coherence tomography angiography by siamese U-Net

Optical coherence tomography angiography (OCTA), as a functional imaging based on OCT, has found successful medical applications. OCTA produces vasculature imaging using blood flow motion as an intrinsic contrast agent. To date, the prevailing OCTA algorithms require B-scan repetitions at the same location, which increases the image acquisition time and eventually the occurrence of motion artifacts. Recently, deep learning models have been developed to improve OCTA reconstruction quality with the same or less repeated B-scans against conventional OCTA algorithms. However, most of the deep learning models feed each repeated B-scan into an individual input channel while not mining the image difference across parallel inputs, which is crucial for creating OCTA image. In this study, we propose a new deep learning model by incorporating Siamese architecture and a new channel fusion method-multi-branch coordinates cross attention module into U-Net, called SU-Net, which is aimed at helping the network to learn the inter-channel difference associated with blood flow better. The SU-Net was evaluated on<i>in vivo</i>OCTA datasets of human skin, including 12 three-dimensional volumetric images from 8 volunteers. Additionally, the generalization of Siamese architecture was investigated on several common deep learning networks, i.e., DnCNN, ResNet, and DenseNet. The results demonstrated that incorporating Siamese architecture can improve deep learning-based OCTA B-scan reconstruction with an improvement of 1.7% to 3.8% in PSNR and 1.2% to 6.4% in SSIM Among all the four networks with Siamese architecture, SU-Net achieved the best OCTA B-scan reconstruction with a PSNR of 28.01 and SSIM of 0.813. In addition, SU-Net yielded a comparable OCTA reconstruction with two images input as compared to U-Net with three images input, leading to a 1.5 times reduction of image acquisition. The results indicate a good potential of our method for clinical applications.