Unet based hybrid quantum-classical generative adversarial network for medical image denoising

Abstract

Medical images are highly susceptible to noise during the image acquisition process that will adversely affect the diagnosis accuracy. There are lots of classical methods available for denoising but they often struggle with the complex and high noise and over smooth the anatomical structures cause loss of fine details and textures. This study focuses whether the employment of quantum computing principle and on generative adversarial network(GAN) can overcome these limitations. The methodology introduces variational quantum circuit (VQC) at the bottleneck layer of the Unet based GAN architecture. The bottleneck VQC uses 16 quibits to transform the classical bottleneck data in quantum states with the strongly entangling layer. VQC transform the compressed latent feature into high dimension Hilbert space enabling the generator compete with PatchGAN discriminator, trained on 10,000 Computed Tomography (CT) images distorted with Additive White Gaussian Noise (AWGN). Quantitative results shows structural similarity index of 0.8987 and peak signal-to-noise ratio of 32.71 dB. These findings significantly outperform the classical filtering. This study shows proposed method outperformed the UNet-GAN classical counterpart and other baseline methods such as DnCNN, BM3D and total variational (TV). Also training is stable and do not mode collapse which generally seen in full classical GAN. Visually quantum-enhanced method produce the subtle details and textures than classical counterparts.