Related Experiment Videos
MuSTAF: Clinically Relevant Multi-task Spatiotemporal Attention Fusion Framework for Breast Cancer Detection with
Yutong Li1, Austin Castelo1, Jennifer B Dennison1
1Y. Li, A. Castelo, K. Brock, and C. Wu are with Imaging Physics; J. B. Dennison and N. M. Kettner are with Clinical Cancer Prevention; W. Sieh is with Epidemiology; O. O. Weaver and C. Wu are with Breast Imaging; and C. Wu is also with Biostatistics, all at The University of Texas MD Anderson Cancer Center, Houston, TX 77030 USA. K. Brock and C. Wu are also with the Institute for Data Science in Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX 77030 USA. J. R. Joseph and E. Castillo are with Biomedical Engineering, The University of Texas at Austin, Austin, TX 78712 USA. C. Wu is also with the Oden Institute for Computational Engineering and Sciences, The University of Texas at Austin, Austin, TX 78712 USA.
A new AI model, MuSTAF, improves breast cancer detection in mammograms by mimicking radiologists' reasoning. It analyzes multiple images over time, enhancing accuracy and providing valuable auxiliary information for clinical decisions.
Area of Science:
- Artificial Intelligence in Medical Imaging
- Breast Cancer Screening
- Radiomics and Computational Pathology
Background:
- Current AI models for mammography often fail to align with clinical reasoning, potentially increasing radiologist workload.
- Radiologists routinely integrate temporal and spatial information, including prior mammograms, breast density, and symmetry, for accurate cancer detection.
Purpose of the Study:
- To develop and evaluate MuSTAF, a multi-task spatiotemporal attention fusion model, for patient-level breast cancer classification using longitudinal mammography.
- To align AI-driven detection with clinical workflows by incorporating temporal dynamics and auxiliary assessments like breast density and laterality.
Main Methods:
- MuSTAF utilizes up to three recent mammograms, integrating temporal and cross-view information with refined suspicious-region features.
- The model performs joint prediction of cancer status, breast density, and bilateral symmetry, with a separate laterality classifier for positive cases.
- Evaluated on internal (n=351) and external (n=8,723) datasets, assessing performance based on cancer classification AUC and auxiliary task performance.
Main Results:
- MuSTAF achieved superior cancer classification AUC (0.84 internal, 0.88 external) compared to baselines and existing models.
- Auxiliary tasks (density/laterality) improved cancer detection performance (AUCs 0.83/0.80).
- Limiting analysis to recent exams (within 60 days) improved external performance (0.72 to 0.88), and three exams outperformed five, highlighting the importance of recent imaging.
Conclusions:
- MuSTAF effectively enhances longitudinal mammographic cancer classification by integrating multi-temporal data and mimicking clinical reasoning.
- The model's ability to provide auxiliary outputs demonstrates its potential to support, rather than burden, clinical decision-making.
- Recent imaging evidence is more critical than remote history for AI-based breast cancer detection, informing optimal AI application in screening.