Zero-Shot Domain Adaptation of 2D Mammography Lesion Detection Model for 3D Breast CT Malignancy Detection
Cancers · 11.08.2026 · 18 (16) · 2585
DOI: 10.3390/cancers18162585 ↗Kurz gefasst
Fachartikel in der Zeitschrift Cancers, online erschienen am 11.08.2026. Verfasst von Andrew P Leynes und 4 weiteren Autor:innen. Beteiligte Einrichtung: Diagnostic and Interventional Radiology, University Hospital Zurich, University Zurich, 8091 Zurich, Switzerland. Die Arbeit nennt 30 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Swiss Cancer League.
Abstract (Originalsprache)
Purpose
Breast cancer screening relies primarily on two-dimensional (2D) digital mammography, which is limited by tissue superposition and compression-related patient discomfort. Dedicated spiral breast computed tomography (BCT) overcomes these limitations by providing isotropic, fully three-dimensional (3D) volumetric images without compression. However, translating artificial intelligence (AI) advancements to 3D BCT lesion detection is hindered by the scarcity of large-scale annotated 3D datasets and substantial computational constraints. The purpose of this study is to evaluate the feasibility of repurposing a pre-trained 2D mammography AI model for 3D BCT malignancy detection without dedicated model retraining.
Methods
In this exploratory proof-of-concept study, the dataset comprised BCT examinations of female patients with malignant lesions. We developed a slab projection framework that systematically rotates the BCT volume and extracts overlapping tissue slabs to simulate mammography-like 2D images compatible with the input requirements of the pre-existing 2D mammography model. The 2D inferences were subsequently back-projected to their original 3D spatial coordinates and aggregated via a consensus voting scheme. Optimization and performance evaluation were performed on the same dataset to establish an internal baseline. Object-level detection performance was assessed using the F1 score, and segmentation quality was evaluated using the Dice coefficient.
Results
Application of the proposed framework enabled 3D lesion localization in BCT volumes without requiring domain-specific training. An optimal configuration was found with a sparser angular sampling combined with medium slab thickness (45° angular spacing, 50-voxel thickness), which achieved an F1 score of 0.346, with 48.3% precision and 26.9% recall. Evaluation of segmentation quality corroborated these findings, yielding a maximum mean Dice coefficient of 0.336 for the 45°/30-voxel configuration.
Conclusions
This exploratory proof-of-concept demonstrates the technical feasibility of adapting a pre-trained 2D mammography AI model to volumetric BCT without architectural modifications or retraining. The reported performance represents an internal same-dataset baseline under substantial domain shift and should not be interpreted as generalizable or clinically deployable performance. These findings provide a foundation for future domain-adapted models and independent external validation of AI-based tumour detection in breast CT.
Schlagwörter
Stichwörter der Autor:innen
Autor:innen und Institutionen
- Diagnostic and Interventional Radiology, University Hospital Zurich, University Zurich, 8091 Zurich, Switzerland.
- b-rayZ, 8952 Schlieren, Switzerland.
- Institute of Diagnostic and Interventional Radiology, GZO Regional Health Center, 8620 Wetzikon, Switzerland.
ORCID-Profile: Jonathan Andreas Saenger, Thomas Frauenfelder
Förderung
- Swiss Cancer League
Zitieren
APA
Leynes, A. P., Saenger, J. A., Weber, A., Frauenfelder, T., & Boss, A. (2026). Zero-Shot Domain Adaptation of 2D Mammography Lesion Detection Model for 3D Breast CT Malignancy Detection. Cancers, 18(16), 2585. https://doi.org/10.3390/cancers18162585
Vancouver
Leynes AP, Saenger JA, Weber A, Frauenfelder T, Boss A. Zero-Shot Domain Adaptation of 2D Mammography Lesion Detection Model for 3D Breast CT Malignancy Detection. Cancers. 2026;18(16):2585. doi:10.3390/cancers18162585
BibTeX
@article{Leynes2026,
author = {Leynes, Andrew P and Saenger, Jonathan Andreas and Weber, Anna and Frauenfelder, Thomas and Boss, Andreas},
title = {Zero-Shot Domain Adaptation of 2D Mammography Lesion Detection Model for 3D Breast CT Malignancy Detection},
journal = {Cancers},
year = {2026},
volume = {18},
number = {16},
pages = {2585},
doi = {10.3390/cancers18162585}
}
EUVIMED