Abstract
Landmark detection remains challenging in large images or when the anatomy is ambiguous. We propose a landmark-first multi-task framework that incorporates anatomical priors through two auxiliary losses: a structure segmentation task providing anatomical context, and a geometric confinement loss regularizing landmark probability maps relative to anatomical shapes. The method is evaluated on two datasets with paired landmark and segmentation annotations: CTPel and ImageTBAD. We compare four objectives with nnLandmark: baseline, segmentation context, geometric confinement, and full (context and confinement). The full objective achieves the best overall localization on CTPel, with a median [inter-quartile range] Euclidean error of 1.5 [0.86] voxels. On ImageTBAD, the performance is comparable across the tested objectives, with the full objective achieving a the best median error of 3.8 [2.2] voxels at a landmark radius of 10 voxels. Across the CTPel radius experiments, the full objective remains the most consistent configuration, especially at smaller landmark radii (p<0.001). For ImageTBAD, geometric confinement significantly improved performance at small radii over the baseline (p<0.001), but the auxiliary structure segmentation task proved ineffective due to the already abundant anatomical context. These results indicate that context and anatomy-aware geometric regularization complement standard segmentation losses for robust landmark detection. The code is available at: https://github.com/LorisGiordano/GeometricConfinement
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ShapeMI_006.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=fKh7AgT8Ne
BibTex
@InProceedings{GioLor_Landmark_MICCAISAT2026,
author = { Giordano, Loris AND Amador Sanchez, Sebastian AND Lenaerts, Tom AND Vandemeulebroucke, Jef},
title = { { Landmark Detection Supported by Anatomy Segmentation and Geometric Constraints } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17259},
month = {pending},
page = {pending}
}