Inpainting non-anatomical objects in brain imaging using enhanced deep convolutional autoencoder network

Abstract

Medical diagnosis can be severely hindered by distorted medical images, especially in the analysis of Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) images. Therefore, enhancing the accuracy of diagnostic imaging and inpainting damaged areas are essential for medical diagnosis. Over the past decade, image inpainting techniques have advanced due to deep learning and multimedia information. In this paper, we proposed a deep convolutional autoencoder network with improved parameters as a robust method for inpainting non-anatomical objects in MRI and CT images. Traditional approaches based on the exemplar methods are much less effective than deep learning methods in capturing high-level features. However, the inpainted regions would appear blurr and with global inconsistency. To handle the fuzzy problem, we enhanced the network model by introducing skip connections between mirrored layers in the encoder and decoder stacks. This allowed the generative process of the inpainting region to directly use the low-level feature information of the processed image. To provide both pixel-accurate and local-global contents consistency, the proposed model is trained with a combination of the typical pixel-wise reconstruction loss and two adversarial losses, which makes the inpainted output seem more realistic and consistent with its surrounding contexts. As a result, the proposed approach is much faster than existing methods while providing unprecedented qualitative and quantitative evaluation with a high inpainting inception score of 10.58, peak signal-to-noise ratio (PSNR) 52.44, structural similarity index (SSIM) 0.95, universal image quality index (UQI) 0.96, and mean squared error (MSE) 40.73 for CT and MRI images. This offers a promising avenue for enhancing image fidelity, potentially advancing clinical decision-making and patient care in neuroimaging practice.