A self-supervised learning-based method for simultaneous motion artifacts removal and denoising of OCTA images

Motion artifacts and noise are key factors that determine image quality in optical coherence tomography angiography (OCTA). Although deep learning has emerged as an effective method for artifacts removal and denoising, its generalization capability remains limited, and it is difficult to handle images with both motion artifacts and noise. To address this issue, we designed a Swin Transformer-based multi-scale motion artifacts and noise parallel removal network (ST-MANPR) to learn the nonlinear mapping between images with motion artifacts and noise and images without them, thereby achieving simultaneous suppression of motion artifacts and noise. The proposed network integrates the Swin window attention module, channel attention module (CAM), and adaptive multi-scale convolution denoising module (AMS-CNN) to enhance its capability in processing complex image features. At the same time, a hybrid loss function combining wavelet transform (WT) and mean square error (MSE) was introduced to facilitate high-frequency detail restoration. In addition, we constructed a dataset of OCTA images with and without motion artifacts and noise at multiple intensity levels. The created dataset was applied to train the network and the test results were evaluated both visually and numerically. The experimental results show that the proposed network can effectively remove motion artifacts and noise in OCTA images simultaneously.