Zero-Shot Domain Adaptation of 2D Mammography Lesion Detection Model for 3D Breast CT Malignancy Detection
Cancers, 2026
Abstract
Introduction: Deploying AI diagnostic tools across different imaging modalities typically demands extensive retraining on large datasets — a resource-intensive process that limits scalability in clinical settings. This study investigated whether b-rayZ's b-box plus, a commercially available AI model originally designed for 2D mammography lesion detection, could be repurposed for malignancy detection in 3D breast CT without any additional model training.
Methods: The researchers applied a zero-shot domain adaptation approach, meaning the 2D mammography AI was transferred directly to analyse 3D breast CT data without retraining or fine-tuning on CT-specific examples. The performance of the adapted model was evaluated for its ability to identify malignant lesions in 3D breast CT scans, assessing how well the model generalised across the significant technical differences between these two imaging modalities.
Results: The zero-shot adapted model demonstrated meaningful malignancy detection capability in 3D breast CT, suggesting that core feature representations learned from 2D mammography data retain clinical utility when applied to volumetric CT imaging. This cross-modality transfer was achieved without incurring the costs and delays associated with curating new training datasets or retraining the AI system.
Conclusion: These findings indicate that existing 2D mammography AI tools may be adaptable to emerging 3D breast imaging workflows without the need for full retraining pipelines. For radiology departments considering investment in breast CT programmes, this approach could substantially reduce the time and cost barriers to deploying validated AI support in new imaging environments.
