Abstract
Automated coronary artery segmentation from X-ray coronary angiography (XCA) remains limited by scarce and heterogeneous annotations. Building on anatomy-guided masked autoencoder (MAE) pretraining, we present a systematic ablation study across three axes: the degree of anatomy-guided masking, the pretraining objective (reconstruction only, reconstruction with a vascular topology-preservation loss, and a further feature-consistency objective), and the fine-tuning strategy (full, partial, frozen). The encoder is pretrained on over 65,000 unlabelled XCA frames and transferred to downstream segmentation using 1,200 labelled images from ARCADE. We evaluate in-domain performance on a held-out test set and out-of-domain performance on an external dataset (DCA1), showing that pretraining consistently improves segmentation metrics over supervised baseline. Notably, full loss pretraining with partial fine-tuning and low anatomy-guided masking achieved the strongest balanced in-domain performance, reaching mean Dice of 0.77 and mclDice of 0.85 on in-domain dataset. Cross-domain analysis showed fine-tuning depth benefited both domains, whereas increasing pretraining objective complexity helped only in-domain at low-tomoderate masking and degraded both domains at high masking. Overall, our results show that MAE-based pretraining improves coronary segmentation robustness, with reconstruction-only pretraining, moderate anatomy-guided masking, and full or partial fine-tuning consistently yielding the most robust transfer in and out of domain datasets.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/TIA_012.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{GutMir_AnatomyAware_MICCAISAT2026,
author = { Gutiérrez-Fernández, Miriam AND Toumpas, Achilleas AND Maga-Nteve, Christoniki AND López-Linares Román, Karen},
title = { { Anatomy-Aware MAE Pretraining for Generalizable Coronary 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}
}