PaperlayerView exact revision on arXiv

Source-linked AI summary

Explicitly Aligned Multi-View Breast Cancer Risk Prediction from Variable-Length Screening Histories

Solveig Thrun, Zijun Sun, Suaiba A. Salahuddin, Kristoffer Wickstrøm, Elisabeth Wetzer, Stine Hansen, Robert Jenssen, Michael Kampffmeyer

arXiv:2610.16329v1cs.CVcs.AI

TL;DR

Breast cancer risk prediction from mammography could use temporal tissue changes and complementary CC and MLO information, but existing methods leave gaps in explicit alignment, joint view modeling, or variable screening histories. FlexLMV-Net aligns and fuses longitudinal multi-view features while accommodating variable prior exams; it shows robust results across datasets and subgroups, and the study introduces BCRBench for standardized comparison.

  • Problem

    Existing explicit-alignment methods model CC and MLO views independently, and explicit longitudinal approaches typically assume a fixed number of prior examinations.

    Source: §1 · p. 2

  • Method

    FlexLMV-Net explicitly aligns prior and current feature maps for both CC and MLO views, fuses them across views over time, and accommodates variable numbers of prior examinations.

    Source: §1 · p. 3

  • Results

    FlexLMV-Net demonstrates robust performance across breast density, cancer subtypes, and demographic subgroups on EMBED and CSAW-CC.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The conclusion reports strong performance across both datasets and diverse subgroups, with additional longitudinal history further improving risk prediction.

    Source: §7 · p. 23

    Further validation on independent multi-center and prospective cohorts is needed to assess generalizability across populations, imaging protocols, and clinical settings.

    Source: §6 · p. 22

Abstract

from arXiv · show

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.

1. Introduction

Age-based screening does not account well for individual risk, while existing longitudinal methods leave gaps in explicit spatial alignment, joint CC/MLO analysis, or variable screening histories. FlexLMV-Net addresses these gaps, and the study also introduces BCRBench for standardized model comparison.

1. Introduction

  • Age-based screening intervals do not adequately account for individual risk, while population screening produces many negative examinations and may underscreen women at high risk.Personalized screening strategies have the potential to improve early detection, reduce screening-related harms, and optimize healthcare-resource use.Source: §1 · p. 2
  • Existing explicit-alignment methods model mammographic views independently, and longitudinal approaches typically assume a fixed number of prior examinations.These constraints limit joint use of complementary CC/MLO information and applicability to patients with different screening histories, including first-time attendees.Source: §1 · p. 2
  • FlexLMV-Net jointly models CC and MLO views by explicitly aligning longitudinal feature maps and integrating aligned representations through cross-view temporal fusion.Its dual-stream attention block uses parallel self- and cross-attention to model intra- and inter-view dependencies.Source: §1 · p. 3
  • FlexLMV-Net is validated on EMBED and CSAW-CC, with reported robust performance across breast density, cancer subtypes, and demographic subgroups.The study also assesses generalization across datasets and clinically relevant subgroups.Source: §1 · pp. 3–4
  • BCRBench is introduced as an open-source benchmark that standardizes evaluation of single- and multi-timepoint risk-prediction models for reproducible comparison.The extended study also expands the methodological description and systematically analyzes component contributions.Source: §1 · p. 3
  • Feature-aware temporal fusion accommodates variable numbers of prior examinations, including no prior examination, broadening the model's applicability to first-time screening participants.The paper describes this as an extension of LMV-Net in the extended study.Source: §1 · p. 3

2. Related Work

Prior work moves from single-timepoint prediction toward longitudinal modeling and richer multi-view fusion, but gaps remain in explicit spatial alignment across views and support for variable histories. FlexLMV-Net is positioned as addressing these gaps in a unified framework.

2.1. Breast Cancer Risk Prediction

  • Single-timepoint mammography models learn predictive imaging biomarkers but do not capture temporal tissue changes; longitudinal models have reported improved risk stratification over single-timepoint approaches.Examples include PRIME+, VMRA-MaR, LoMaR, and TRINet; these methods primarily rely on implicit alignment.Source: §2.1 · p. 4
  • Implicit alignment is sensitive to geometric variation across visits, while explicit spatial registration has been reported to better capture longitudinal changes and improve risk prediction.OA-BreaCR learns feature-space deformation fields, whereas ImgFeatAlign estimates image-based deformation fields and applies them to feature representations.Source: §2.1 · p. 4
  • Existing explicit-alignment approaches model views independently, and typically fixed prior-exam counts constrain their use across differing screening histories.FlexLMV-Net is described as combining explicit alignment, joint CC/MLO modeling, and variable numbers of prior examinations.Source: §2.1 · pp. 4–5

2.2. Multi-view mammography analysis

  • Multi-view architectures have progressed from late fusion of CC/MLO branches to spatial self- and cross-view interaction and prompt-tuning approaches.MV-Swin-T reports outperforming multi-view CNN baselines across benchmark datasets; MVPT-NET integrates views using visual prompt tuning.Source: §2.2 · p. 5
  • Longitudinal multi-view models use GRU branches, visit aggregation, or side encoding and temporal attention to combine information across views and examinations.LRP-NET targets near-term risk estimation; LoMaR aggregates multiple visits and all four standard views; VMRA-MaR models asymmetric cues.Source: §2.2 · p. 5
  • Existing multi-view longitudinal models either lack explicit registration across timepoints or assume fixed prior-exam counts; FlexLMV-Net combines alignment, cross-view temporal fusion, and variable input handling.The gap is jointly using spatially aligned multi-view information across patients with differing screening histories.Source: §2.2 · p. 5

3. Methodology

FlexLMV-Net models five-year breast cancer risk by aligning longitudinal CC and MLO features, fusing variable-length histories, and integrating complementary views. Its risk head predicts from fused representations, with view-specific supervision during training.

3. Methodology

  • The framework aligns prior and current mammograms, fuses longitudinal information, models cross-view dependencies, and predicts five-year risk from fused CC and MLO representations.View-specific features provide auxiliary supervision during training.Source: §3 · p. 5

3.1. Explicit Longitudinal Multi-View Alignment

  • Image-derived deformation fields align each available prior feature map to the current examination in both CC and MLO views, establishing spatial correspondence for variable-length histories.A shared encoder processes the views, and aligned features support temporal-change modeling.Source: §3.1 · p. 6

3.2. Feature-Aware Temporal Fusion

  • Feature-Aware Temporal Fusion independently aggregates aligned prior and difference features, weighting examinations by feature content, temporal distance, and spatial relevance.The fused features are combined with current features, while screening-time gaps are incorporated to handle variable histories and intervals.Source: §3.2 · pp. 7–8, §3.1 · p. 7

3.3. Multi-View Dual-Stream Attention

  • Parallel self-attention and cross-attention streams capture within-view spatial dependencies and corresponding information across CC and MLO views.The stream outputs are fused and refined with residual connections to produce updated feature maps.Source: §3.3 · p. 9

3.4. Risk Prediction Head

  • The risk head globally pools and fuses refined view features, then estimates cumulative risk across time with auxiliary view-specific predictions during training.A masking function restricts loss contributions to observed, clinically valid time points in censored data.Source: §3.4 · pp. 10–11

4. Experimental Setup and BCRBench

The experimental setup evaluates longitudinal risk prediction on EMBED and CSAW-CC using standardized preprocessing, five-year risk metrics, and a common benchmarking framework.

4.1.1. Datasets

  • EMBED contains 364,564 mammograms from 23,256 patients, while CSAW-CC includes 873 cancer cases and 7,850 controls.EMBED uses images from three vendors; CSAW-CC uses Hologic images, and the datasets have eight- and seven-year negative follow-up periods, respectively.Source: §4.1.1 · p. 12

4.1.2. Time-to-Events label generation

  • EMBED positives are identified using BI-RADS 6 or pathology, whereas CSAW-CC positives include screen-detected and interval cancers labeled using radiological timing.CSAW-CC cancer timing is inferred from the most recent examination because biopsy dates are unavailable; only negative prior screening mammograms are included.Source: §4.1.2 · p. 12

4.1.3. Preprocessing

  • Images are converted from DICOM to standardized 16-bit PNGs, segmented to retain breast regions, resized to 1664 × 2048, and intensity-normalized.EMBED density uses BI-RADS categories; CSAW-CC density estimates are grouped into low, medium, and high categories.Source: §4.1.3 · p. 12

4.2. Evaluation Metrics

  • Models are assessed on one- to five-year risk prediction using C-index and AUC, with 1,000-resample bootstrap confidence intervals and paired tests for improvement.Improvements over the second-best model are considered significant at p < 0.05.Source: §4.2 · p. 13

4.3. Implementation Details

  • FlexLMV-Net is compared with Mirai, VMRA-MaR, OA-BreaCR, ImgFeatAlign, and LMV-Net; it uses a fine-tuned Mirai encoder and up to three prior examinations.Models are trained for up to 30 epochs, with early stopping; the three-prior limit reflects memory constraints.Source: §4.3 · p. 13

4.4. BCRBench

  • BCRBench unifies preprocessing, data handling, training, and evaluation to support reproducible comparisons of breast cancer risk prediction models.Its modular interface includes EMBED and CSAW-CC, single-timepoint and longitudinal models, and subgroup evaluation using AUC and C-index.Source: §4.4 · p. 13

5. Results

The results compare LMV-Net and FlexLMV-Net with prior methods, then assess performance across breast density and cancer-type subgroups. The longitudinal models generally perform strongly, with FlexLMV-Net achieving the strongest overall C-index and subgroup results.

5.1. Comparison with State-of-the-Arts (SOTA)

  • FlexLMV-Net achieves the strongest overall C-index, while LMV-Net and FlexLMV-Net consistently outperform competing methods across EMBED and CSAW-CC.FlexLMV-Net improves on LMV-Net at several horizons; improvements over the best baseline are statistically significant (p < 0.05).Source: §5.1 · pp. 14–15

5.2. Performance across density categories

  • Across density categories, both longitudinal models maintain strong performance on EMBED and CSAW-CC, with FlexLMV-Net particularly strong for short-term prediction.On EMBED, FlexLMV-Net is comparable to or better than LMV-Net in the highest-density categories; differences are smaller on CSAW-CC.Source: §5.2 · p. 15, Figure 6 · p. 15, Figure 7 · p. 16

5.3. Performance across cancer types

  • FlexLMV-Net achieves the highest C-index across cancer-type subgroups on both datasets, including CSAW-CC tumor-size strata.The comparison with LMV-Net shows improvement when more prior examinations are incorporated.Source: §5.3 · p. 15, Figure 9 · p. 17

5.4. Performance across demographic subgroups

The paper evaluates demographic subgroup performance, cross-dataset generalization, and attention maps. FlexLMV-Net performs well across demographic groups, while the generalization results qualify the benefit of additional history under domain shift.

5.4. Performance across demographic subgroups

  • On EMBED, longitudinal models generally exceed SOTA C-index across race groups, and FlexLMV-Net is highest or near-highest.Model differences are more pronounced for African American or Black and Asian groups than for Caucasian or White participants.Source: §5.4 · p. 17

5.5. Generalization performance

  • FlexLMV-Net improves in-domain performance over LMV-Net, but additional history does not consistently improve cross-dataset performance under domain shift.LMV-Net maintains the strongest cross-dataset performance in some settings; EMBED-trained models generalize more consistently to CSAW-CC than in the reverse direction.Source: §5.5 · pp. 17–18, Table 2 · p. 18

5.6. Attention map analysis

  • Attention maps focus on suspicious and tumor regions, with cross-attention generally showing higher Attention Density Ratio than self-attention across four examples.Tumors are shown across left and right breasts; attention density within tumor regions is 8–12× that in surrounding tissue.Source: §5.6 · pp. 18, 21, Figure 13 · p. 20

5.7. Effect of longitudinal history length on risk prediction

The results assess how screening-history length and model components relate to risk prediction performance. A limited history improves prediction, and the full FlexLMV-Net configuration performs best in the ablation study.

5.7. Effect of longitudinal history length on risk prediction

  • Using prior mammograms consistently improves prediction over no prior images, with performance largely plateauing after two or three examinations.The benefit is particularly pronounced at longer follow-up horizons, where performance declines substantially without prior imaging.Source: §5.7 · p. 21

5.8. Ablation study

  • The full FlexLMV-Net performs best on both datasets; removing longitudinal information, multi-view modeling, dual-stream attention, or encoder fine-tuning reduces performance.Using one prior examination also performs worse than the full configuration using up to three prior examinations.Source: §5.8 · pp. 21–22, Table 3 · p. 22

6. Limitations and future works

The authors identify limits in validation, alignment robustness, screening-history length and selection, bilateral information, and imaging modality.

Validation

  • Independent multicenter and prospective cohorts are needed to assess generalizability across populations, imaging protocols, and clinical settings.Source: §6 · p. 22

Alignment

  • Alignment errors may propagate to risk prediction because explicit feature-space alignment depends on accurate spatial correspondence.Source: §6 · p. 22

Screening history

  • Future work could investigate longer screening histories and strategies for selecting and weighting prior examinations.Source: §6 · p. 22

Bilateral information

  • Bilateral breast information may provide complementary risk-related signals through between-breast asymmetries.Source: §6 · p. 22

Imaging modality

  • Extending the framework to digital breast tomosynthesis could leverage three-dimensional structure and assess its complementary value alongside conventional mammography.Source: §6 · p. 22

7. Conclusion

The conclusion presents FlexLMV-Net as a variable-history longitudinal multi-view framework and LMV-Net as its one-prior special case. It reports strong performance across two datasets and diverse subgroups, and introduces BCRBench for standardized comparison.

Conclusion

  • FlexLMV-Net supports variable numbers of prior examinations through explicit longitudinal feature alignment and feature-aware temporal fusion of CC and MLO views.Source: §7 · pp. 22–23
  • LMV-Net is a fixed-length special case of FlexLMV-Net with K = 1 prior examination.Source: §7 · p. 23
  • Both frameworks demonstrated strong performance across EMBED and CSAW-CC and diverse subgroups, while additional longitudinal history further improved risk prediction.Source: §7 · p. 23
  • BCRBench unifies breast cancer risk prediction models in a standardized framework to facilitate fair and reproducible comparisons.Source: §7 · p. 23

CRediT authorship contribution statement

The contribution statement assigns conceptualization and methodology across the author team, with distinct software, analysis, supervision, funding, and project-administration roles.

Authorship contributions

Appendix A. Dataset Details and Dataset Splits

The appendix summarizes time-to-event labels, dataset splits, and breast density distributions for EMBED and CSAW-CC, with a density-data caveat for EMBED.

Dataset details

Breast density

  • For EMBED, breast density labels are unavailable for some cases.Source: Appendix A · p. 24
  • The number of EMBED cases in the density distribution table therefore differs from the total in the time-to-cancer table.Source: Appendix A · p. 24

CSAW-CC

Table A.5 reports breast-density distributions for breast-level CC+MLO exam pairs in EMBED and CSAW-CC, with percentages calculated within each split.

CSAW-CC

  • Table A.5 covers breast-density distributions across EMBED and CSAW-CC for breast-level CC+MLO exam pairs.Source: Table A.5 · p. 24
  • Percentages are reported relative to the total number of cases in each split.Source: Table A.5 · p. 24

Related papers

Loading 2610.16329v1…