Optical coherence tomography (OCT) has the characteristics of high resolution, high sensitivity and high speed. However, affected by factors such as the high scattering of tissue, the micro-movement of target, and the jitter of hardware during imaging process, OCT images always carry noise dominated by speckle noise, which reduces the accuracy of the subsequent processing. How to denoise the image to improve the image quality has been highly concerned. Current denoising methods based on deep learning are almost end-to-end, which means that the denoising degree is uncontrollable. However, the noise intensity may be different in different cases, and its uncontrollable denoising degree will lead to the reduction of the generalization ability of the model. For doctors, the denoising degree required is different depending on the patient's condition. The end-to-end deep learning network limits the autonomous control of the denoising degree. Therefore, achieving end-to-process denoising is of great significance in clinical applications.

The TMI_2013OCT dataset publicly available from Duke University is used in this work, which is obtained from the normal population and patients with age-related macular degeneration (AMD). In order to avoid the under-fitting problem of the model caused by insufficient training data, data augmentation is used to expand the size of training set to 79200 pairs. Using multi-layer convolution and deconvolution to build an autoencoder, a modularized denoising autoencoder (MDAE) is built based on the architecture of a modular deep neural network. Each autoencoder module can sequentially output a image with gradually increased denoising degree. Process results meet different usage requirements. In order to reduce the amount of parameters and improve the training speed, all modules share parameters (i.e., all modules have the same parameters). Mean square error (MSE), peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are used as evaluation metrics.

To quantitatively evaluate the denoising capability with different number of modules (T), the MSE, PSNR and SSIM are calculated when T varies from 1 to 4 (Fig. 4). The MSE has dropped significantly after the denoising of the first module, and then is still decreased but the magnitude is getting smaller and smaller after the denoising of the subsequent each module. Both PSNR and SSIM have been greatly improved after denoising in the first module, and the magnitude gradually becomes smaller after subsequent each module. These metrics show that the proposed end-to-process model can denoise the image progressively, and T=4 is the best choice in this work after considering factors such as denoising performance and running time. To evaluate the performance of the proposed model, retinal OCT images of normal eyes are randomly selected from the test set for testing and compared with Gaussian filtering, mean filtering, block-matching and 3D filtering (BM3D), and stacked denoising autoencoder (SDAE) methods (Fig. 5). On the premise of maintaining a significant denoising effect, the proposed method preserves more image details, and has the best results in all the metrics compared to others (Table 1). To further examine the denoising performance of the proposed method on retinal OCT images of diseased eyes, images from AMD patients are randomly selected from the test set for testing (Fig. 6). After MDAE denoising, there is almost no information loss, and the image restoration is the highest. All the metrics of MDAE except time are the best among all methods (Table 2), indicating that the proposed method also has the best performance for retinal OCT images of diseased eyes. Gaussian filtering and mean filtering methods have absolute time advantages with very little calculation, but the denoising effect is very unsatisfactory. The BM3D method is the most time-consuming, and the average time to process an image is close to 7.0 s, which is unbearable in practical clinical applications. The average time for MDAE method to process each image is about 0.26 s. Although it has no advantage compared to other methods and is far from the requirement of real-time denoising, denoising usually belongs to the post-processing stage of a image, and this time consumption is still at an acceptable level.

Compared to the original retinal OCT image without processing, the PSNR of the denoised result obtained by the proposed method is increased by 11.32 dB and 12.08 dB for the images from normal eyes and diseased eyes, respectively, and the noise level is greatly reduced. This provides the possibility for subsequent high-precision image processing and analysis. The proposed method can control the denoising degree by controlling the number of modules, so as to be more suitable for complex clinical applications. At the same time, the proposed method only relies on the parameter T to control the denoising degree without adjusting other parameters, which saves the user's learning cost and is very friendly to doctors who focus on clinical tasks.