Adversarial Distortion Learning for Medical Image Denoising

We present a novel adversarial distortion learning (ADL) for denoising two-and three-dimensional (2D/3D) biomedical image data. The proposed ADL consistsof two auto-encoders: a denoiser and a discriminator. The denoiser removesnoise from input data and the discriminator compares the denoised result to itsnoise-free counterpart. This process is repeated until the discriminator cannotdifferentiate the denoised data from the reference. Both the denoiser and thediscriminator are built upon a proposed auto-encoder called Efficient-Unet.Efficient-Unet has a light architecture that uses the residual blocks and anovel pyramidal approach in the backbone to efficiently extract and re-usefeature maps. During training, the textural information and contrast arecontrolled by two novel loss functions. The architecture of Efficient-Unetallows generalizing the proposed method to any sort of biomedical data. The 2Dversion of our network was trained on ImageNet and tested on biomedicaldatasets whose distribution is completely different from ImageNet; so, there isno need for re-training. Experimental results carried out on magnetic resonanceimaging (MRI), dermatoscopy, electron microscopy and X-ray datasets show thatthe proposed method achieved the best on each benchmark. Our implementation andpre-trained models are available at https://github.com/mogvision/ADL.