Accelerating Prostate DWI Scans: A Combination of Local SVD and Deep Learning for Enhanced Denoising

Motivation: Diffusion-weighted MR images often suffer from low signal-to-noise ratio, particularly at high b-values, diminishing their diagnostic value. To counter this, multiple repetitions per diffusion direction are typically acquired and averaged, which is time-consuming and prone to motion artifacts. Goal(s): Present a method that reduces the required number of repetitions in DWI, thus shortening scan times, while preserving diagnostic value. Approach: The repetitions in DWI are jointly denoised through a combination of local Singular Value Decomposition and deep-learning-based denoising. Results: Our evaluations indicate that this approach outperforms competing methods, offering a potential solution to the problem of prolonged acquisition times in DWI. Impact: By combining local singular value decomposition with deep-learning-based denoising techniques, the necessary number of repetitions for the acquisition of diffusion-weighted MR images is substantially decreased and thus the acquisition is accelerated, while retaining comparable image quality.