Medical image reconstruction using compressible latent space invertible networks

Many modern medical imaging systems are computed in nature and need computational reconstruction to form an image. When the acquired measurements are insufficient to uniquely determine the sought-after object, prior knowledge about the object needs to be utilized in order to successfully recover an image. An emerging line of research involves the use of deep generative models such as generative adversarial networks (GANs) as priors in the image reconstruction procedure. However, when GANs are employed, reconstruction of images outside the range of the GAN leads to errors that result in realistic but false image estimates. To address this issue, an image reconstruction framework is proposed that is formulated in the latent space of an invertible generative model. A novel regularization strategy is introduced that takes advantage of the architecture of certain invertible neural networks (INNs). To evaluate the performance of the proposed method, numerical studies of reconstructing images from stylized MRI measurements are performed. The method outperforms classical reconstruction methods in terms of traditional image quality metrics and is comparable to a state-of- the-art adaptive GAN based framework, while benefiting from a deterministic procedure and easier tuning of the regularization parameters.