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

Marco Nittscher, Michael Falk Lameter, Riccardo Barbano, Johannes Leuschner, Bangti Jin, Peter Maass
Medical Imaging with Deep Learning, PMLR 227:617-642, 2024.

Abstract

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

Cite this Paper

BibTeX
@InProceedings{pmlr-v227-nittscher24a, title = {SVD-DIP: Overcoming the Overfitting Problem in DIP-based CT Reconstruction}, author = {Nittscher, Marco and Lameter, Michael Falk and Barbano, Riccardo and Leuschner, Johannes and Jin, Bangti and Maass, Peter}, booktitle = {Medical Imaging with Deep Learning}, pages = {617--642}, year = {2024}, editor = {Oguz, Ipek and Noble, Jack and Li, Xiaoxiao and Styner, Martin and Baumgartner, Christian and Rusu, Mirabela and Heinmann, Tobias and Kontos, Despina and Landman, Bennett and Dawant, Benoit}, volume = {227}, series = {Proceedings of Machine Learning Research}, month = {10--12 Jul}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v227/nittscher24a/nittscher24a.pdf}, url = {https://proceedings.mlr.press/v227/nittscher24a.html}, abstract = {The deep image prior (DIP) is a well-established unsupervised deep learning method for image reconstruction; yet it is far from being flawless. The DIP overfits to noise if not early stopped, or optimized via a regularized objective. We build on the regularized fine-tuning of a pretrained DIP, by adopting a novel strategy that restricts the learning to the adaptation of singular values. The proposed SVD-DIP uses ad hoc convolutional layers whose pretrained parameters are decomposed via the singular value decomposition. Optimizing the DIP then solely consists in the fine-tuning of the singular values, while keeping the left and right singular vectors fixed. We thoroughly validate the proposed method on real-measured μCT data of a lotus root as well as two medical datasets (LoDoPaB and Mayo). We report significantly improved stability of the DIP optimization, by overcoming the overfitting to noise.} }
Endnote
%0 Conference Paper %T SVD-DIP: Overcoming the Overfitting Problem in DIP-based CT Reconstruction %A Marco Nittscher %A Michael Falk Lameter %A Riccardo Barbano %A Johannes Leuschner %A Bangti Jin %A Peter Maass %B Medical Imaging with Deep Learning %C Proceedings of Machine Learning Research %D 2024 %E Ipek Oguz %E Jack Noble %E Xiaoxiao Li %E Martin Styner %E Christian Baumgartner %E Mirabela Rusu %E Tobias Heinmann %E Despina Kontos %E Bennett Landman %E Benoit Dawant %F pmlr-v227-nittscher24a %I PMLR %P 617--642 %U https://proceedings.mlr.press/v227/nittscher24a.html %V 227 %X The deep image prior (DIP) is a well-established unsupervised deep learning method for image reconstruction; yet it is far from being flawless. The DIP overfits to noise if not early stopped, or optimized via a regularized objective. We build on the regularized fine-tuning of a pretrained DIP, by adopting a novel strategy that restricts the learning to the adaptation of singular values. The proposed SVD-DIP uses ad hoc convolutional layers whose pretrained parameters are decomposed via the singular value decomposition. Optimizing the DIP then solely consists in the fine-tuning of the singular values, while keeping the left and right singular vectors fixed. We thoroughly validate the proposed method on real-measured μCT data of a lotus root as well as two medical datasets (LoDoPaB and Mayo). We report significantly improved stability of the DIP optimization, by overcoming the overfitting to noise.
APA
Nittscher, M., Lameter, M.F., Barbano, R., Leuschner, J., Jin, B. & Maass, P.. (2024). SVD-DIP: Overcoming the Overfitting Problem in DIP-based CT Reconstruction. Medical Imaging with Deep Learning, in Proceedings of Machine Learning Research 227:617-642 Available from https://proceedings.mlr.press/v227/nittscher24a.html.

Related Material