Fast OCT deconvolution using a light-weight CNN

The broad application of deconvolution in OCT is hindered by speckle-noise-induced deconvolution artifacts and the time-consuming deconvolution process. To address these issues, we propose a deep-learning-based method for fast OCT deconvolution, which include a speckle noise reduction module and a deconvolution module. We also introduce a lightweight convolutional neural network (CNN) for accelerating the inference processes. With similar performance on artifact-free deconvolution, our method is 2,777 times faster than the state-of-the-art iterative deconvolution algorithm and achieves an average inference time of 1.41ms for a single B-frame. The number of parameters of our proposed CNN architecture is 2.17 times less than that of U-Net, and the inference is 5 times faster. For the discrete objects in OCT images, our method achieves 50.58% and 60.36% improvements in axial and transverse resolutions, respectively. For the continuous objects, our method achieves an average 13.91dB improvements in the contrast-to-noise ratio.