Abstract
Unsupervised representation learning is widely used to characterize anatomical variability in medical imaging, yet its robustness under multi-center domain shift remains unclear, particularly for complex anatomical structures such as the human ear, which exhibits highly intricate geometry and subtle inter-subject variability. In this work, we investigate whether learned embeddings capture meaningful anatomical traits or are instead dominated by acquisition-specific biases in ear CT data. We compare image-based embeddings derived from a dual-task 3D U-Net with shape-based latent codes learned via a neural unsigned distance field auto-decoder, using simple landmark-derived geometric features as interpretable proxies of anatomical variability. Experiments are conducted across three heterogeneous datasets under both single-dataset and multi-dataset training regimes, and evaluated using retrieval metrics and latent space analyses. Results show that image-based embeddings are strongly influenced by dataset characteristics, with clustering primarily reflecting acquisition properties rather than underlying ear anatomy. While multi-dataset training reduces this effect, domain bias persists. Importantly, shape-based embeddings do not fully mitigate this behavior, indicating that even geometry-focused representations of this complex anatomical structure can inherit domain effects. These findings highlight the need for careful validation of learned representations and demonstrate that simple, interpretable anatomical features provide robust baselines under domain shift.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/iMIMIC_016.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: Not Available
BibTex
@InProceedings{BelSar_Anatomical_MICCAISAT2026,
author = { Belinchón Castro, Sara AND Jiménez Ordoñez, Aïda AND López Diez, Paula},
title = { { Anatomical Information or Domain Shortcuts? Probing Image- and Shape-Based Embeddings in Ear CT with Landmark-Derived Features } },
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}
}