Abstract
Deep learning methods for 3D shape generation and mesh-based reconstruction have advanced rapidly; however, the relation between geometry and learned latent representations remains poorly understood, particularly for shapes. Existing explainable AI methods for 3D shapes predominantly rely on gradient-based techniques, are architecture-specific, or depend on end-to-end differentiability. These limitations challenge their implementation in non-differentiable pipelines of complex generative models. We propose RESHAPE, a model-agnostic framework for latent representation explainability that does not require access to gradients or internal model structure. RESHAPE estimates the geometric importance of individual latent code components by applying stochastic masking to the latent codes, reconstructing the perturbed shapes, and statistically quantifying the geometric deviation of the resulting reconstructions from the original. We demonstrate the properties of RESHAPE on a controlled synthetic dataset and validate its practical relevance on medical shape modeling tasks. Our framework enables (i) highlighting latent dimensions with strong geometric influence within the latent spaces of models, (ii) generating latent saliency maps, and (iii) separating primary shape-defining from stochastic latent dimensions.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/iMIMIC_007.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/iMIMIC_007_supp.pdf
Link to Open Review
Open Review Page: Not Available
BibTex
@InProceedings{AlRam_RESHAPE_MICCAISAT2026,
author = { Al-Belmpeisi, Rami AND Wickstrøm, Kristoffer Knutsen AND Sundgaard, Josefine Vilsbøll AND Larsen, Peter Hjørringgaard AND Dahl, Anders Bjorholm AND Jenssen, Robert AND Kampffmeyer, Michael C.},
title = { { RESHAPE: Representation Learning for the Explainability of Shapes } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17279},
month = {pending},
page = {pending}
}