BIOMEDICAL IMAGING GROUP (BIG)
Laboratoire d'imagerie biomédicale (LIB)

Neuronal Architectures with Trainable Activations for Stable Medical Image Reconstruction

M. Unser

Keynote address, Inverse Problems and Artificial Intelligence in Medicine (IPAIM'25), Bath, United Kingdom, June 23-July 4, 2025.

The two dominant CNN-based paradigms for consistent image reconstruction are: (i) the use of trainable regularizers, and (ii) proximal-gradient-type architectures (PnP) with a learned denoiser. To ensure stability, these approaches require constraints—either by imposing (weak) convexity on the regularizer or by enforcing that the trained denoiser in PnP be 1-Lipschitzs—which severely limits the expressivity of such systems. We show that this limitation can be mitigated by making the neuronal activations trainable as well, subject to appropriate slope constraints. This is achieved by imposing a second-order total variation penalty on each trainable activation, which results in adaptive linear spline solutions. We then show how the deep spline framework enables the efficient training of such systems. We illustrate the potential of our approach through denoising and biomedical image reconstruction experiments, and report promising results. 

@INPROCEEDINGS(http://bigwww.epfl.ch/publications/unser2503.html,
AUTHOR="Unser, M.",
TITLE="Neuronal Architectures with Trainable Activations for Stable
	Medical Image Reconstruction",
BOOKTITLE="Inverse Problems and Artificial Intelligence in Medicine
	({IPAIM'25})",
YEAR="2025",
editor="",
volume="",
series="",
pages="",
address="Bath, United Kingdom",
month="June 23-July 4,",
organization="",
publisher="",
note="Keynote address")
© 2025 IPAIM. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from IPAIM. This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.