EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC 4.0

AbdomenAtlasNet-3.0: Multi-Phase Deep Learning for Robust Multi-Organ Segmentation of Core Abdominal Structures on a Large-Scale CT Atlas

Sasan Shafiei, Danial Soltani, Moein Aboobakri Makooei, Amirreza Geranfar, Mahshid Rezaei, Iman Razipour

InfoScience Trends · 2025

DOI: 10.61882/ist.202502.10.03 ↗

Kurz gefasst

Fachartikel in der Zeitschrift InfoScience Trends, 2025. Verfasst von Sasan Shafiei und 5 weiteren Autor:innen. Die Arbeit nennt 34 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Accurate multi-organ segmentation on abdominal CT scans is essential for quantitative analysis, treatment planning, and clinical decision support. However, segmentation performance remains limited for small or low-contrast organs, and existing deep learning models rarely exploit the complementary information provided by multi-phase abdominal CT. This study aims to develop and evaluate an attention-based multi-phase fusion architecture capable of robustly segmenting core abdominal organs across large-scale, heterogeneous datasets. Using the AbdomenAtlas 3.0 dataset—comprising more than 9,000 abdominal CT volumes from multiple institutions—we identified 3,220 scans with at least one contrast-enhanced phase and complete annotations for nine core abdominal structures. We proposed Abdo-menAtlasNet-3.0, a multi-encoder, attention-driven intermediate fusion network that integrates non-contrast, arterial, and portal-venous CT phases while accommodating missing phases via modality-presence masks. Performance was compared with a single-phase baseline (SP-UNet) and an early-fusion model (EF-UNet) using Dice similarity coefficient (DSC), Hausdorff distance (HD95), and statistical testing on an institution-stratified test set. AbdomenAtlasNet-3.0 achieved superior segmentation accuracy across all organs, with a macro-averaged DSC of 0.921 ± 0.004 compared with EF-UNet (0.906 ± 0.005) and SP-UNet (0.893 ± 0.006). Improvements were most pronounced for small or anatomically challenging organs—including the pancreas, gallbladder, and inferior vena cava—showing DSC gains of 3–6 points over baseline models. The proposed method demonstrated consistent cross-institution robustness and maintained competitive performance even in single-phase studies. Attention-based intermediate fusion of multi-phase CT significantly enhances abdominal multi-organ segmentation, particularly for difficult structures, and generalizes effectively across diverse scanners and institutions. AbdomenAtlasNet-3.0 establishes a scalable and phase-aware framework for reliable abdominal segmentation and provides a foundation for future multimodal or downstream clinical AI applications.

Zitieren

APA

Shafiei, S., Soltani, D., Makooei, M. A., Geranfar, A., Rezaei, M., & Razipour, I. (2025). AbdomenAtlasNet-3.0: Multi-Phase Deep Learning for Robust Multi-Organ Segmentation of Core Abdominal Structures on a Large-Scale CT Atlas. InfoScience Trends. https://doi.org/10.61882/ist.202502.10.03

Vancouver

Shafiei S, Soltani D, Makooei MA, Geranfar A, Rezaei M, Razipour I. AbdomenAtlasNet-3.0: Multi-Phase Deep Learning for Robust Multi-Organ Segmentation of Core Abdominal Structures on a Large-Scale CT Atlas. InfoScience Trends. 2025. doi:10.61882/ist.202502.10.03

BibTeX
@article{Shafiei2025,
  author = {Shafiei, Sasan and Soltani, Danial and Makooei, Moein Aboobakri and Geranfar, Amirreza and Rezaei, Mahshid and Razipour, Iman},
  title = {AbdomenAtlasNet-3.0: Multi-Phase Deep Learning for Robust Multi-Organ Segmentation of Core Abdominal Structures on a Large-Scale CT Atlas},
  journal = {InfoScience Trends},
  year = {2025},
  doi = {10.61882/ist.202502.10.03}
}
Kostenlos starten →