Abstract
Synthesis of novel anatomical shapes via deep generative models provides a promising route for applications such as virtual imaging trials, but extending these methods to anatomically plausible collections of organs remains challenging. In particular, overlap between adjacent organ shapes violates anatomical plausibility. Here, we introduce a differentiable, SDF-based loss term that penalizes overlap and can be used during supervised learning, as well as for gradient-based latent optimization during inference, providing weak and strong guarantees, respectively.
We demonstrate our approach in the joint synthesis of the thyroid and the surrounding organs, including the trachea, cervical esophagus, and common carotid arteries.
We find that while the baseline model accurately reconstructs anatomy and generates novel, diverse cases, most samples exhibit inter-organ overlap.
Regularization substantially reduces initial overlap at negligible cost with respect to reconstruction accuracy, while post-hoc optimization eliminates this overlap in few iterations without affecting novelty or diversity.
Combining both yields completely overlap-free anatomy in only one to two post-hoc iterations on average.
An observer study further suggests that the synthesized anatomy is difficult to distinguish from real anatomy, supporting its plausibility.
Source code is available at \href{https://github.com/MIAGroupUT/overlap-free-shape-synthesis}{https://github.com/MIAGroupUT/overlap-free-shape-synthesis}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/Off_Grid_011.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=CY3kWc8Ep7
BibTex
@InProceedings{deBra_OverlapFree_MICCAISAT2026,
author = { de Wilde, Bram AND Rietberg, Max T. AND Dima, Alina F. AND van Aalst, Joëlle E. AND Lajoinie, Guillaume AND Wolterink, Jelmer M.},
title = { { Overlap-Free Multi-Organ Shape Synthesis with Implicit Neural Representations } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17277},
month = {pending},
page = {pending}
}