SVD-DIP: Overcoming the Overfitting Problem in DIP-based CT Reconstruction

• DOI: 10.48550/arxiv.2303.15748
• Publication Date: 2023-01-01
• Authors: Marco Nittscher; Michael Lameter; Riccardo Barbano; Johannes Leuschner; Bangti Jin; Peter Maaß
• Keywords: Overfitting; Singular value

The deep image prior (DIP) is a well-established unsupervised learning method for reconstruction; yet it far from being flawless. DIP overfits to noise if not early stopped, or optimized via regularized objective. We build on the fine-tuning of pretrained DIP, by adopting novel strategy that restricts adaptation singular values. proposed SVD-DIP uses ad hoc convolutional layers whose parameters are decomposed value decomposition. Optimizing then solely consists in values, while keeping left and right vectors fixed. thoroughly validate real-measured $\mu$CT data lotus root as well two medical datasets (LoDoPaB Mayo). report significantly improved stability optimization, overcoming overfitting noise.