What Does Anatomical Shape Know About You? a Multi-Organ, Shape-Only Study of Demographic Prediction, Attribution, and Privacy

MCML Authors

Christian Wachinger

Prof. Dr.

Core PI

Artificial Intelligence in Medical Imaging

Abstract

How much does the bare shape of internal organs reveal about a person once all image intensity is removed? Using the public TotalSegmentator CT and MRI datasets (1,228 whole-body CTs), we convert the provided segmentations of 16 anatomical structures into surface point clouds and study, in one controlled setting, demographic prediction, per-organ attribution, size versus scale-free shape, shape versus intensity-based features, cross-site and cross-modality generalization, and privacy. Within this single cohort, shape alone predicts sex at AUC 0.866 (bootstrap 95% CI 0.846–0.886) and age within 9.3 years; the hip bones are the strongest sex cue (AUC 0.940). Shape outperforms a Hounsfield intensity baseline for sex (0.868 vs. 0.729), so against this baseline the sex signal is primarily geometric rather than a density artifact. This signal persists under leave-one-institution-out evaluation (AUC 0.827) and replicates on MRI (AUC 0.762); a classifier trained on CT shapes transfers to MRI without retraining (AUC 0.708). Much of the sex signal is organ size, yet scale-free shape alone still reaches AUC 0.791. The same signal has a constructive side: as a calibrated auto-labeling tool, it sexes the most confident half of subjects at 94% accuracy. It also carries privacy risk. A single organ’s quantized shape signature makes 93% of subjects unique within the cohort, rising with more organs; this distinctiveness also depends on how finely the descriptors are binned (18% unique at 3 bins, 99% at 20). We report these as upper bounds on linkability, not re-identification rates. Even so, 'de-identified' organ meshes remain highly distinctive and disclose demographics.

inproceedings LHE+26

---

ShapeMI @MICCAI 2026

Workshop on Shape in Medical Imaging at the 29th International Conference on Medical Image Computing and Computer Assisted Intervention. Strasbourg, France, Sep 27-Oct 01, 2026. To be published. Preprint available.

---

Authors

G. Luijten • B. Hinrichs-Puladi • M. Engelke • C. Wachinger • J. Egger

---

Links

URL GitHub

---

Research Area

 C1 | Medicine