Explicitly Aligned Multi-View Breast Cancer Risk Prediction from Variable-Length Screening Histories
Authors: Solveig Thrun, Zijun Sun, Suaiba A. Salahuddin, Kristoffer Wickstrøm, Elisabeth Wetzer, Stine Hansen, Robert Jenssen, Michael Kampffmeyer
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 · Norwegian Computing Center, Oslo, 0373, Norway · Pioneer Centre for AI, University of Copenhagen, Copenhagen, 1350, Denmark
Accurate breast cancer risk prediction from screening mammography is essential for improved risk stratification and personalized screening. It can benefit from both temporal breast tissue changes across longitudinal examinations and complementary anatomical information from craniocaudal (CC) and mediolateral oblique (MLO) views. Leveraging longitudinal information requires identifying corresponding tissue regions across examinations, yet most approaches rely on implicit feature alignment without explicitly enforcing spatial consistency. Recent work has demonstrated the benefits of explicit temporal alignment; however, existing methods model each view independently, preventing them from exploiting complementary CC and MLO information. Furthermore, explicit approaches have been limited to a single prior examination. To fill these gaps, we propose (github.com/sot176/FlexLMV-Net), a breast-wise longitudinal multi-view framework that jointly models CC and MLO views across screening examinations while accommodating a variable number of prior examinations. FlexLMV-Net explicitly aligns longitudinal feature maps and integrates complementary views through cross-view temporal fusion and a dual-stream attention block. Evaluated on EMBED and CSAW-CC, FlexLMV-Net consistently outperforms existing approaches across breast density, cancer subtypes, and demographic subgroups, highlighting the benefits of explicit longitudinal alignment, multi-view modeling, and flexible use of available screening history. We further introduce (github.com/sot176/BCRBench), an open-source benchmark that unifies state-of-the-art single- and multi-timepoint breast cancer risk prediction models under standardized preprocessing, training, inference, and evaluation protocols, enabling fair and reproducible comparison of existing methods.
Figures & tables
Figure 1: Overview of the proposed FlexLMV-Net architecture. The framework consists of three main components: (1) explicit longitudinal multi-view alignment to establish correspondence between current and prior examinations, (2) a multi-view dual-stream attention block that models intra-view and inter-view dependencies across longitudinal data, and (3) a risk prediction head that leverages fused multi-view representations, with auxiliary view-specific supervision during training, to predict five-year breast cancer risk.
Figure 2: Overview of the explicit longitudinal multi-view feature alignment in FlexLMV-Net. For illustration, the alignment process is shown for the MLO view, where prior feature maps are aligned to the current features using image-derived deformation fields, enabling consistent correspondence and the construction of view-specific longitudinal representations. The CC view is processed in exactly the same manner.
Figure 3: The Feature-aware temporal fusion of FlexLMV-Net adaptively weights longitudinal features based on temporal importance and spatial relevance.
Figure 4: Overview of the Multi-View Dual-Stream Attention block in FlexLMV-Net. Longitudinal features from the MLO and CC views are tokenized and processed through parallel self- and cross-attention streams to model intra- and inter-view dependencies. The resulting features are then refined via residual connections and reshaped back into spatial feature maps, yielding enhanced multi-view representations.
Figure 5: Risk prediction head of FlexLMV-Net. Refined MLO and CC features are globally pooled and fused for risk estimation via a cumulative probability layer, with auxiliary view-specific branches providing additional supervision during training.
Figure 6: C-index across breast density categories for the proposed LMV-Net and state-of-the-art methods on the EMBED and CSAW-CC datasets (in-domain evaluation). Points show the mean C-index and error bars represent the 95% confidence interval.
Figure 7: Forest plots showing AUC across five follow-up years for all models stratified by breast density categories (BI-RADS A-D) on the EMBED dataset. Points represent mean AUC and error bars represent 95% confidence intervals.
Figure 8: Forest plots showing AUC across five follow-up years for all models stratified by breast density categories (Low, Medium, High) on the CSAW-CC dataset. Points represent mean AUC and error bars represent 95% confidence intervals.
Figure 9: Forest plots showing C-index for all models stratified by cancer types on the EMBED and CSAW-CC dataset. Points represent the mean C-index, and error bars indicate 95% confidence intervals.
Figure 10: C-index values with 95% confidence intervals for different race categories on the EMBED dataset.
Figure 11: Generalization performance of models across (a) density categories and (b) cancer subtypes. Models are evaluated both in-domain (trained and tested on the same dataset) and cross-dataset (trained on one dataset and tested on the other) to assess generalization. Points show mean C-index and error bars indicate 95% confidence intervals.
Figure 12: C-index across race categories for models evaluated on the EMBED dataset. Models were trained either on EMBED (in-domain) or CSAW-CC (cross-dataset). Points represent the mean C-index, and error bars indicate the 95% confidence interval.
Figure 13: Attention maps across multiple cases showing that FlexLMV-Net correctly focuses on tumor regions (highlighted by red boxes).
Figure 14: Effect of longitudinal imaging history on breast cancer risk prediction. AUC is shown across 1-, 2-, and 3-year follow-up horizons using up to three prior mammographic examinations.
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 · Department of Computer Science and Engineering, University of Bologna, 40126, Bologna, Italy +4