The Impact of Longitudinal Mammogram Alignment on Breast Cancer Risk Assessment
Authors: Solveig Thrun, Stine Hansen, Zijun Sun, Nele Blum, Suaiba A. Salahuddin, Xin Wang, Kristoffer Wickstrøm, Elisabeth Wetzer, +3 more
Organizations: Department of Physics and Technology, UiT The Arctic University of Norway, Tromsø, 9037, Norway · SPKI The Norwegian Centre for Clinical Artificial Intelligence, University Hospital of North Norway, Tromsø, 9019, Norway · Department of Computer Science and Engineering, University of Bologna, 40126, Bologna, Italy · Fraunhofer Research Institution for Individualized and Cell-Based Medical Engineering IMTE, 23562, Lübeck, Germany · Department of Radiology, Netherlands Cancer Institute (NKI), Amsterdam, 1066 CX, The Netherlands · Norwegian Computing Center, Oslo, 0373, Norway · Pioneer Centre for AI, University of Copenhagen, Copenhagen, 1350, Denmark
Regular mammography screening is crucial for early breast cancer detection. By leveraging deep learning-based risk models, screening intervals can be personalized, especially for high-risk individuals. While recent methods increasingly incorporate longitudinal information from prior mammograms, accurate spatial alignment across time points remains a key challenge. Misalignment can obscure meaningful tissue changes and degrade model performance. In this study, we provide insights into various alignment strategies, image-based registration, feature-level (representation space) alignment with and without regularization, and implicit alignment methods, for their effectiveness in longitudinal deep learning-based risk modeling. Using two large-scale mammography datasets, we assess each method across key metrics, including predictive accuracy, precision, recall, and deformation field quality. Our results show that image-based registration consistently outperforms the more recently favored feature-based and implicit approaches across all metrics, enabling more accurate, temporally consistent predictions and generating smooth, anatomically plausible deformation fields. Although regularizing the deformation field improves deformation quality, it reduces the risk prediction performance of feature-level alignment. Applying image-based deformation fields within the feature space yields the best risk prediction performance. These findings underscore the importance of image-based deformation fields for spatial alignment in longitudinal risk modeling, offering improved prediction accuracy and robustness. This approach has strong potential to enhance personalized screening and enable earlier interventions for high-risk individuals. The code is available at https://github.com/sot176/Mammogram_Alignment_Study_Risk_Prediction.git, allowing full reproducibility of the results.
Figures & tables
Figure 1: Overview of longitudinal risk prediction methods: (a) Using only the current mammogram, which serves as our single–timepoint risk model, (b) Direct feature extraction without alignment, (c) Implicit Alignment, (d) Feature-level alignment, (e) Image-level alignment with MammoRegNet, (f) Applying MammoRegNet’s deformation field in feature space, and (g) Risk prediction using alignment methods (d), (e), and (f).
Figure 2: Analysis of C-index by density category for the different alignment methods across both datasets.
Figure 3: AUC-ROC curves for risk predictions on the EMBED dataset across 1-5 year outcomes: comparison of different alignment methods.
Figure 4: Comparison of different alignment methods in predicting time to cancer development. (a) Precision, (b) Recall.
Figure 5: Density-stratified AUC performance of different alignment methods for predicting time to cancer development. (a) Short-term (1-year) risk prediction. (b) Long-term (5-year) risk prediction.
Figure 6: Comparison of model performance for non-invasive and invasive cancer. C-index values with 95% confidence intervals are shown.
Figure 7: AUC for non-invasive cancer for each year with 95% confidence intervals.
Figure 8: AUC for invasive cancer for each year with 95% confidence intervals.
Figure 9: C-index values with 95% confidence intervals for different race categories.
Figure 10: AUC for the different races.
Figure 11: Comparison of deformation field quality. The top row shows displacement vectors, and the bottom row displays Jacobian determinant maps (white/blue: valid; orange/red: invalid non-invertible regions).
Figure 12: Comparison of deformation field quality. The top row shows displacement vectors, and the bottom row displays Jacobian determinant maps (white/blue: valid; orange/red: invalid non-invertible regions).
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 13: Visualization of the architectures for the longitudinal risk prediction.
Figure 14: Architecture of the proposed MammoRegNet.
Figure 15: Registration result of MammoRegNet. Visual results of the MammoRegNet registration performance for patients without cancer. (a) Current mammogram, (b) prior mammogram, (c) aligned prior mammogram, (d) current mammogram overlaid with the prior mammogram, (e) current mammogram overlaid with the affine-transformed prior mammogram, and (f) current mammogram overlaid with the final transformed prior mammogram.
Figure 16: Registration result of MammoRegNet. Visual results of the MammoRegNet registration performance for patients with cancer. The tumor region is highlighted by a bounding box on the current mammogram and the aligned prior mammogram. (a) Current mammogram, (b) prior mammogram, (c) aligned prior mammogram, (d) current mammogram overlaid with the prior mammogram, (e) current mammogram overlaid with the affine-transformed prior mammogram, and (f) current mammogram overlaid with the final transformed prior mammogram.
Department of Physics and Technology, UiT The Arctic University of Norway, Tromsø, Norway · SPKI The Norwegian Centre for Clinical Artificial Intelligence, University Hospital of North Norway, Tromsø, Norway · Norwegian Computing Center, Oslo, Norway +1
Department of Physics and Technology, UiT The Arctic University of Norway, Tromsø, 9037, Norway · SPKI The Norwegian Centre for Clinical Artificial Intelligence, University Hospital of North Norway, Tromsø, 9019, Norway · Norwegian Computing Center, Oslo, 0373, Norway +1