TL;DR for operators

A screening program may have several years of mammograms for the same patient. Using those examinations can improve predictions of future breast cancer risk, but including more historical images does not necessarily produce better results.

FlexLMV-Net1 combines anatomical alignment, time-aware historical-image fusion, and multi-view attention to estimate breast cancer risk over one to five years. It achieves strong retrospective discrimination on two datasets, yet its improvement over a single-prior predecessor is inconsistent: the multiple-prior model improves overall discrimination on CSAW-CC but falls slightly behind on EMBED. The simpler predecessor also performs better in both reported cross-dataset transfer tests.

For hospital AI evaluators and breast screening programs, the result changes the model-selection question. The objective is not to maximize the number of historical examinations processed, but to determine whether those additional inputs improve performance on the institution’s own population enough to justify their retrieval, alignment, and processing requirements. The accompanying BCRBench framework supports standardized comparisons, but clinical outcomes, calibrated intervention thresholds, and economic benefits remain unestablished.

More screening history produces uneven gains

Consider a patient returning for routine breast screening. The institution has the current mammogram and three earlier examinations. A prediction system could use only the most recent comparison or retrieve and process the complete available history.

The additional images might reveal changes that a single comparison misses. They might also introduce differences in positioning, compression, imaging equipment, or acquisition timing. The operational question is whether the model can extract useful longitudinal information without mistaking these differences for meaningful tissue changes.

The researchers evaluate how well models rank future breast cancer risk. Their principal overall measure is the concordance index, or C-index, which assesses agreement between predicted risk ordering and observed event timing. A C-index of 81% does not mean that 81% of patients were correctly diagnosed, nor does it establish that predicted probabilities are calibrated.

The results show a clear difference between the two evaluation datasets.

Evaluation dataset Single-prior LMV-Net Multiple-prior FlexLMV-Net Difference
EMBED 81.4% 81.2% −0.2 pp
CSAW-CC 74.4% 77.1% +2.7 pp

Source: Sections 5.1 and 5.8, Tables 1 and 3. Values are overall C-indices for the fine-tuned configurations. Differences are percentage points.

On CSAW-CC, the multi-prior architecture delivers a meaningful increase in measured discrimination. On EMBED, it does not improve the aggregate result.

A separate history-length experiment, restricted to patients with at least three earlier examinations, finds that performance generally plateaus with two or three priors. The evidence therefore supports using historical information, but not an assumption that its incremental predictive value grows continuously with history length.

The distinction also matters experimentally: the single-prior and multiple-prior configurations differ in their history-processing architecture. Their performance difference should not be interpreted as the isolated effect of supplying one additional image.

How FlexLMV-Net makes historical images comparable

Mammograms taken years apart may show the same tissue under different physical positioning and compression. Directly comparing features at the same image coordinates can therefore confuse anatomical change with acquisition variation.

FlexLMV-Net addresses this problem through three connected operations.

First, a shared Mirai-based image encoder extracts spatial feature representations from current and historical mammograms. A pretrained, frozen MammoRegNet estimates deformation fields, which are used to align historical feature maps with the current examination. The model can then construct temporal differences between more anatomically corresponding representations.

Second, feature-aware temporal fusion assigns importance weights to available historical examinations. These weights depend on current and historical image features and elapsed time. Spatial attention further adjusts their contribution at different image locations.

This means that two prior examinations need not contribute equally, even when both are available. An examination’s value depends on its content, temporal relationship, and spatial information rather than its mere inclusion in the record.

Third, dual-stream attention combines information from the two standard mammographic views: craniocaudal (CC) and mediolateral oblique (MLO). Self-attention processes relationships within each view, while cross-attention allows information from one projection to inform the representation of the other. Auxiliary view-specific prediction heads provide additional supervision during training.

The resulting features feed cumulative one- to five-year risk prediction heads. Training uses follow-up masking so that incompletely observed prediction horizons do not contribute as though their outcomes were known.

Although the architecture accommodates variable numbers of examinations, the implemented experiments use at most three priors because of memory constraints. Its performance with substantially longer histories is not established.

The ablations show why history and attention still matter

The mixed multiple-prior results do not imply that historical images contribute little.

Component-removal experiments in Section 5.8 test what happens when selected parts of the architecture are disabled or replaced.

Model configuration EMBED C-index CSAW-CC C-index
Full FlexLMV-Net 81.2% 77.1%
No prior; current features repeated 77.8% 73.0%
Only one mammographic view 70.8% 73.3%
Without dual-stream attention 72.8% 69.5%

Source: Section 5.8, Table 3. These are architectural ablations, not prospective clinical intervention studies.

Removing historical information reduces discrimination by 3.4 percentage points on EMBED and 4.1 points on CSAW-CC. Removing dual-stream attention produces larger reductions: 8.4 and 7.6 points, respectively.

These comparisons support two conclusions. Historical information contributes useful predictive signal within the tested architecture, and joint attention over complementary views contributes substantially to its measured discrimination.

They do not isolate every underlying anatomical mechanism or prove that the attention weights constitute faithful explanations. The paper’s attention visualizations show concentration around annotated tumors in four illustrative patients, but such examples cannot establish causal interpretability.

The findings also depend on the prediction horizon. On CSAW-CC, FlexLMV-Net achieves a one-year AUC of 82.5%, yet its five-year AUC is 72.3%, below the best evaluated external baseline’s 75.2%.

A higher overall C-index should therefore not substitute for evaluating the specific risk horizon that a screening decision would use.

Cross-dataset testing changes the preferred model

A hospital selecting a risk prediction model faces a different question from a researcher optimizing results on an existing benchmark.

A model trained on one screening population may encounter different imaging equipment, population characteristics, and data distributions elsewhere. Cross-dataset evaluation tests whether its predictive discrimination persists when applied to another dataset without retraining.

The study evaluates both transfer directions between the US-based EMBED dataset and Sweden’s CSAW-CC dataset.

Training → evaluation Single-prior LMV-Net FlexLMV-Net
EMBED → CSAW-CC 72.9% 72.5%
CSAW-CC → EMBED 73.7% 72.9%

Source: Section 5.5, Table 2. Values are cross-dataset C-indices for fine-tuned configurations.

In both directions, the single-prior model performs better.

FlexLMV-Net trained on EMBED falls from an 81.2% in-domain C-index to 72.5% on CSAW-CC. Trained on CSAW-CC, it falls from 77.1% in-domain to 72.9% on EMBED.

These experiments are robustness tests, not evidence identifying the precise cause of performance deterioration. Differences in patient populations, acquisition systems, labels, and other dataset characteristics are not individually isolated.

Nevertheless, the model-selection implication is concrete: the architecture that processes more history and achieves the highest score within one dataset is not necessarily the architecture that transfers most effectively to another screening population.

What hospital evaluation teams can do with BCRBench

The paper also introduces BCRBench, an open-source framework that standardizes preprocessing, training, inference, and evaluation across supported mammography risk prediction architectures.

That standardization addresses a recurring evaluation problem: performance comparisons are difficult to interpret when methods use different data handling, encoder configurations, or experimental conventions.

Cognaptus interpretation: A hospital AI evaluation team can use this kind of shared experimental framework to structure a model-selection process before considering changes to screening practice.

The evaluation should compare competing architectures on locally representative examinations, with results separated by prediction horizon, available history length, imaging context, and relevant patient subgroup. Missing or unusable prior examinations and registration failures are especially relevant operational conditions because FlexLMV-Net relies on aligned longitudinal inputs.

The study also provides a reason to account for additional processing requirements. On the reported eight-GPU training setup, FlexLMV-Net requires 22 hours on EMBED versus 16 hours for fine-tuned LMV-Net, and 26 versus 17 hours on CSAW-CC. These are research training measurements, not established deployment costs or end-to-end workflow latency.

The relevant procurement decision is whether locally demonstrated predictive gains compensate for additional historical-image retrieval, registration, computation, and maintenance requirements.

BCRBench can help make that comparison consistent. It cannot establish that a model is suitable for an institution whose imaging population or operating conditions differ from the research datasets.

Better risk ranking is not yet better screening

The evidence concerns retrospective prediction using 59,306 breast-level examination pairs across EMBED and CSAW-CC. These observations are not 59,306 distinct patients.

The study includes subgroup analyses and cross-dataset evaluation, strengthening its characterization of predictive performance. But its results do not establish equivalent performance across every clinical subgroup, accurate absolute risk probabilities in deployment settings, or improved patient outcomes.

Three decisions consequently remain outside the evidence: which predicted-risk thresholds should trigger additional imaging, how screening intervals should change, and whether such interventions produce net clinical and economic benefit.

Those decisions require appropriate calibration, independent prospective validation, and assessment of downstream consequences. The reported results do not justify automating them.

FlexLMV-Net advances the technical handling of irregular screening histories, particularly through explicit alignment and feature-aware multi-view fusion. Its experiments also demonstrate why architectural sophistication must be evaluated separately from external reliability.

For screening institutions, the central finding is that historical imaging can improve risk discrimination, while the value of each additional examination is conditional on the dataset, model configuration, and evaluation objective. That conditional value—not the maximum amount of history a model can ingest—is what should govern subsequent validation and model selection.

Research artifacts

Reference

Cognaptus: Automate the Present, Incubate the Future.


  1. Solveig Thrun and Zijun Sun and Suaiba A. Salahuddin and Kristoffer Wickstrøm and Elisabeth Wetzer and Stine Hansen and Robert Jenssen and Michael Kampffmeyer (2026). Explicitly Aligned Multi-View Breast Cancer Risk Prediction from Variable-Length Screening Histories. arXiv:2610.16329. https://arxiv.org/abs/2610.16329 ↩︎