Abstract
Brain MR image transformation, which is the process of transforming MR images of one type to another, is critical to several downstream neuroimaging tasks that include brain tissue and lesion volume estimation and lesion detection. In recent years, several deep learning-based methods have been applied to address this task; however, for the most part, they have tended to be deterministic. These methods provide a single transformed output for a given input image with no accompanying measure of confidence in the transformation process. In contrast to this, in this study, we demonstrate how a class of probabilistic conditional generative algorithms can be applied to MR image transformation and quantify the performance of these algorithms. We also demonstrate that the ability to generate multiple transformed images for a given input image can be used to estimate the uncertainty in the output and to detect out-of-distribution (OOD) input images. In particular, we apply conditional Generative Adversarial Networks (cGAN), Noise Conditional Score Networks (NCSN), and Denoising Diffusion Probabilistic Models (DDPM) to transform T1, T2, FLAIR, and proton density (PD) MR images. Through extensive computational experiments, we conclude that the probabilistic algorithms are more accurate than other benchmark methods, and among these, the diffusion models yield the most accurate transformation results. Within the diffusion models, DDPM demonstrates higher performance in terms of similarity metrics, and NCSN exhibits accurate distributional measures and computationally favorable characteristics. We also demonstrate how the generative models can be used to assess the confidence in a given transformation and to detect input images that contain pathology and/or artifacts.