Optical coherence tomography (OCT) is employed as a safe and effective diagnostic tool for a variety of ophthalmic diseases due to its high resolution and non-invasive imaging, which is regarded as the "gold standard" in ophthalmic disease diagnosis. However, various kinds of noise, especially speckle noise, seriously affect the quality of retinal OCT images to reduce the contrast and resolution, which makes it difficult to segment and measure retinal sublayer thickness at the pixel level. Therefore, it is of significance to reduce the noise of retinal OCT images and retain structural details such as layering and edges of the images to the greatest extent. The deep learning-based noise reduction method shows advantages in image quality, especially in preserving edge details. However, for in vivo imaging, it is difficult to obtain a large number of multi-frame registration ground truth images, which affects the performance of the supervised learning method. Therefore, the realization of unsupervised denoising independent of ground truth images is vital in the clinical diagnosis of eye diseases.

We propose an unsupervised deep residual sparse attention network (DRSA-Net) based on the Noise2Noise training strategy for retinal OCT image denoising. DRSA-Net consists of local sparse attention block (LSAB), depth extraction block (DEB), global attention block (GAB), and residual block (RB). The TMI_2013OCT dataset publicly provided by Duke University is selected and preprocessed, and a total of 7800 Clean-Noisy and Noisy-Noisy image pairs are obtained. The proposed DRSA-Net is compared with the classical deep learning denoising networks U-Net and DnCNN from two aspects of qualitative visual evaluation and quantitative numerical evaluation and is also compared with the traditional BM3D algorithm. Then the denoising effects of three convolutional neural networks under supervised learning and unsupervised learning strategies are compared. Finally, generalization ability tests and network module ablation experiments are performed based on another public retinal OCT image dataset.

The results of unsupervised training denoising (Fig. 3) show that the built model has better denoise and intra-layer fine structure preservation ability for retinal OCT images. U-Net-N2N tends to destroy the details and boundary of layers and introduces some fuzzy structures among layers. DnCNN-N2N brings degradation of layer boundary and blurring of the outer limiting membrane. The comparison between the results of supervised training and unsupervised training (Fig. 4) indicates that when ideal ground truth images cannot be provided, the denoised images of the supervised learning model have more noise, while the unsupervised learning model has a higher denoise degree and can provide clearer structures and edge information. The denoising numerical evaluation results of supervised learning and unsupervised learning (Table 1) show that compared with the original noise images, the supervised learning and unsupervised learning models realize great improvement in various evaluation indexes of the images. Additionally, compared with the traditional block matching algorithm BM3D, the denoising algorithm based on deep learning reduces the denoising time by two orders of magnitude. High-quality noise reduction of OCT images can be achieved within 1 s, and the proposed algorithm can get ahead of most evaluation indexes regardless of what kind of training strategy is adopted. The test results of generalization ability of unsupervised learning (Fig. 5) show that our proposed model has better generalization ability among different datasets, and can obtain a cleaner background than ground truth in terms of background denoise. In terms of structural information retention, it has clearer interlayer structures and more uniform layers. The results of ablation experiments on different modules of the denoising network proposed (Table 3) indicate that the combination of LSAB+DEB+GAB+RB is better in various evaluation indexes, which fully demonstrates the contribution of each module in the network structure to high-quality noise reduction.

We put forward an unsupervised depth residual sparse attention denoising algorithm independent of ground truth images to solve the noise interference in retinal OCT images and the difficulty of acquiring high-quality multi-frame average images in in vivo imaging. The attention mechanism is combined with sparse convolution kernel to complete the information mining between data efficiently and fully, and the Noise2Noise training strategy is adopted to complete the high-quality training with noise images, which achieves a high level of noise reduction and preserves the multi-layer structure information of retinal OCT images. The traditional denoising algorithm and the classical deep learning network are compared and analyzed from the visual evaluation and numerical evaluation including PSNR, SSIM, EPI, and ENL respectively. The denoising effect of supervised learning and unsupervised learning and the experimental results of the generalization ability test on the public retinal OCT image dataset show that the proposed noise reduction algorithm yields good results in various evaluation indexes and has strong generalization. Compared with supervised learning, unsupervised learning can still obtain better noise reduction performance under insufficient data sets.