A Regularized Conditional GAN for Posterior Sampling in Inverse Problems
A novel regularization technique applicable for medical imaging applications that leverages conditional generative adversarial networks (cGANs) to generate reconstructed images in significantly shorter timeframes.
Several techniques are used for image reconstruction in the medical arena, such as magnetic resonance imaging (MRI) and computed tomography (CT). Conventional techniques rely on generating a single ‘best’ image, are suboptimal in image quality, and lack sample diversity. This lack of diversity can be traced back to inadequate training capabilities of the different models applied and insufficient calibration of weights. Addressing these issues can help enforce consistency, optimal weights, and automatic calibration resulting in accelerated MR image recovery and faster large-scale image completion/inpainting.
The proposed approach focuses on techniques that enable the exploration of the space of probable images instead of a single ‘best’ image. It advances techniques such as Wasserstein GAN to generate high-quality posterior samples at a high rate, building upon proven research. It infuses quantitative metrics to improve image quality and assist the physician in diagnosis. The method uses enhanced generator and discriminator architectures to allow the samples generated to agree regarding means and covariance attributes with the true posterior images while producing samples at a significantly higher rate than conventional approaches.
Medical imaging (Multicoil MR reconstruction, large-scale image inpainting) could benefit from the proposed approach for improved performance – better images, a larger selection of images, and enhanced diagnostic capability. Other applications such as CT, super-resolution, and de-blurring would also show gains in image generation and other downstream activities, such as making the correct diagnosis based on the images generated.
The technique outperforms state-of-the-art GAN and other competing methods in terms of rate/speed of image generation (i.e., 4000x faster), accuracy metrics, and perpetual metrics, e.g., CFID, PSNR, and SSIM.