Abstract
Accurate segmentation of pulmonary tubular structures is essential for lung surgical planning and quantitative pulmonary analysis, yet remains challenging due to distal branch sensitivity, complex geometries, and low-intensity contrast. Although recent deep learning methods improve accuracy, they often fail to preserve structural integrity and rely on heavy architectures or extensive annotations. Designing loss functions that explicitly encode structural priors offers a promising alternative. We propose Cupido (Consistent Unified Prior for Intricate Directional Optimization), a plug-and-play loss for thin, complex tubular structures. Cupido integrates two constraints: a directional consistency term enforcing local angular smoothness, and a geometry-aware union term unifying skeleton- and distance-based priors to enhance global coherence. We evaluate Cupido with three backbones (nnUNet, UMambaBot, and UxLSTMEnc) on two public pulmonary segmentation benchmarks comprising three evaluation subsets, training all models under identical settings for 50 epochs to evaluate robustness across backbones and datasets. Experiments show consistent improvements across backbones and datasets. On the ATM’22 Batch2 dataset, Cupido achieves 91.59% Dice, 86.38% NSD, and 86.80% clDice with UxLSTMEnc, exceeding the baseline by +1.78%, +2.75%, and +3.26%, respectively. It improves fine-branch delineation and structural consistency. By encoding directional and geometric priors, Cupido enables robust, anatomically consistent predictions with minimal training cost for clinically relevant tubular anatomy analysis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/TIA_002.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{YuwCha_Cupido_MICCAISAT2026,
author = { Yuwen, Chang AND Yang, Litao AND Mahapatra, Dwarikanath AND Bozorgtabar, Behzad AND Ge, Zongyuan},
title = { { Cupido Loss for Structure-Aware Pulmonary Airway and Artery Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17269},
month = {pending},
page = {pending}
}