.png){kind=icon}
.png){kind=icon}
Return
Multi-scale feature fusion and reweighting based domain generalization for mammogram classification
.png){kind=icon}
PRE
AI
Y
J
W
W
Z
C
刘
DOI：10.1007/s00530-026-02572-8.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Early detection of breast cancer, followed by timely intervention, plays a crucial role in improving patient survival rates. Mammogram, as a unique diagnostic method for breast cancer, is the preferred choice for early screening. Deep learning-based computer-aided diagnostic tools have shown methodological potential for mammogram analysis. However, there are scale differences in the images captured by different scanners. Models trained on data captured by one scanner often perform poorly on data captured by another scanner. The lack of generalization ability acts as a significant barrier to deploying models in the constantly evolving clinical environment. Enhance the model’s generalizability across different mammogram scanners. We propose a domain generalization framework for mammogram classification that incorporates multi-scale feature fusion and feature reweighting mechanisms (MSFFR). MSFFR includes a feature extractor, a multi-scale feature fusion module, and a multi-scale feature reweighting module. Additionally, the multi-scale feature fusion module enhances the model’s understanding of pathological structures through multi-scale information interaction, reducing reliance on device style information. By adjusting the importance of different channels, the multi-scale feature reweighting module promotes the learning of disease-relevant domain-invariant features. Experimental results on the public INbreast dataset and two private datasets (InH1 and InH2) demonstrate that the proposed MSFFR method achieves superior performance compared with several benchmark approaches. Compared to other domain generalization methods, our approach shows stronger cross-domain generalization for mammogram classification in retrospective experiments.
Keywords:
Domain generalization
Classification
Mammography
Computer-aided diagnosis
Journal
IF:
3.1
Papers:
2.7K
Citations:
2.7K
Organization
Cited Papers