SpurBreast: Controlling Spurious Correlations in Breast MRI

• SpurBreast is a curated breast MRI dataset that systematically encodes spurious correlations as controllable experimental variables.
• It couples image data with over 100 supplementary clinical and acquisition features to analyze both global and local spurious signals.
• Benchmark splits show that models exploiting spurious cues achieve high training accuracy yet fail on unbiased tests, highlighting generalization challenges.

Searching arXiv for the SpurBreast paper and closely related work on spurious correlations in medical imaging. arXiv search query: title:"SpurBreast" OR abs:"spurious correlations" AND abs:"breast MRI" SpurBreast is a curated breast MRI dataset introduced to investigate spurious correlations in real-world medical image classification. It is designed around a specific failure mode of deep neural networks: the ability to learn non-clinical features instead of medically meaningful patterns, thereby achieving high apparent performance under biased sampling while failing to generalize to unbiased data. The dataset is derived from the DUKE Breast Cancer Dataset and combines image data with extensive supplementary metadata, while also providing controlled benchmark splits that either intentionally incorporate or explicitly remove selected spurious correlations. Within this framework, two dominant spurious signals are emphasized: magnetic field strength, treated as a global feature, and image orientation, treated as a local feature (Won et al., 2 Oct 2025).

1. Dataset provenance and composition

SpurBreast is derived from the DUKE Breast Cancer Dataset, which contains 3D MRI scans from over 900 patients with biopsy-confirmed invasive breast cancer (Won et al., 2 Oct 2025). Each patient has a 3D scan with an average of 250 slices. For the SpurBreast benchmark, only slices distinctly categorized as tumor-positive or tumor-negative are retained, while slices in a buffer zone between tumor and non-tumor regions are discarded (Won et al., 2 Oct 2025).

The imaging modality is 3D MRI acquired using 1.5T or 3T scanners. In addition to the image volumes, SpurBreast includes rich tabular data for over 100 demographic, clinical, and imaging device or protocol features, including ethnicity, menopause status, surgical details, and imaging device parameters (Won et al., 2 Oct 2025).

This construction makes the dataset unusual within medical imaging benchmarks. Rather than supplying images alone, it couples image evidence with a broad feature space that can be examined for statistical dependence with the predictive label. This suggests that SpurBreast is intended not merely as a classification corpus, but as an experimental substrate for studying the interaction between image-level prediction and auxiliary patient, device, and acquisition variables.

2. Metadata structure and curation logic

A central property of SpurBreast is the explicit use of supplementary variables during dataset construction. Over 100 supplementary features were analyzed for correlations with the predictive label, with features selected based on their presence, label availability, and balance (Won et al., 2 Oct 2025). The resulting benchmark therefore depends on a prior feature-screening stage rather than on incidental dataset bias alone.

The dataset organizes information at multiple levels: image content, demographic variables, and protocol or device descriptors (Won et al., 2 Oct 2025). In practical terms, this enables the same underlying breast MRI material to be partitioned in qualitatively different ways. A split can be constructed so that the target label is statistically independent of a candidate nuisance factor, or instead so that the nuisance factor is perfectly aligned with the label in training and validation data (Won et al., 2 Oct 2025).

This curation logic is methodologically important because it converts spurious correlation from an uncontrolled observational artifact into a controllable experimental variable. Existing medical imaging datasets, as described in the paper, are not designed to systematically study this issue, largely due to restrictive licensing and limited supplementary patient data (Won et al., 2 Oct 2025). SpurBreast therefore occupies a distinct position: it operationalizes shortcut learning as a benchmark condition rather than treating it solely as a post hoc explanation for poor external validity.

3. Deliberately encoded spurious signals

SpurBreast identifies two dominant spurious signals. The first is magnetic field strength, defined by acquisition on 1.5T versus 3T scanners and described as a global feature influencing the entire image through resolution, artifacts, and contrast (Won et al., 2 Oct 2025). The second is image orientation, implemented as vertical flipping of MR slices via data augmentation and described as a local feature affecting spatial alignment without changing the underlying tissue (Won et al., 2 Oct 2025).

The paper also reports other investigated but less robust spurious factors due to data limitations, including ethnicity, menopause status, and surgery type (Won et al., 2 Oct 2025). However, magnetic field strength and image orientation are the primary benchmarked signals because they were identified as the strongest and most experimentally tractable.

The distinction between global and local spurious features is analytically useful. Magnetic field strength can alter image-wide appearance statistics, potentially allowing a model to classify based on acquisition style rather than pathology. Image orientation, by contrast, changes spatial arrangement while leaving tissue semantics intact, permitting direct study of sensitivity to non-clinical spatial regularities. This suggests that SpurBreast is structured to probe shortcut learning across more than one scale of nuisance variation.

4. Benchmark split construction and evaluation protocol

The benchmark uses a fixed, unbiased test set comprising 150 randomly chosen patients with corresponding slices isolated for all evaluations (Won et al., 2 Oct 2025). For training and validation under spurious conditions, patients are split so that the target label is perfectly aligned to a supplementary feature. In the magnetic field experiment, for example, all tumor slices in train and validation come from 1.5T scans and all non-tumor slices from 3T. In the orientation experiment, only one orientation is linked to the positive label (Won et al., 2 Oct 2025).

Alongside these biased splits, SpurBreast provides baseline datasets without spurious correlations, where random patient splits ensure independence between the label and any specific feature (Won et al., 2 Oct 2025). The benchmark therefore contains both shortcut-favoring and shortcut-neutral conditions.

The reported experimental setup uses ResNet-50 and ViT-B/16, both pretrained on ImageNet (Won et al., 2 Oct 2025). Optimization is selected through grid search on learning rate, optimizer choice among SGD, Adam, and AdamW, and weight decay. Models are trained for 50 epochs, and the best model is selected via validation accuracy (Won et al., 2 Oct 2025). The paper reports Accuracy, Positive Predictive Value, and Negative Predictive Value, and repeats key experiments for magnetic field strength and orientation 10 times to confirm robustness to sampling (Won et al., 2 Oct 2025).

The benchmark logic is direct: if a model exploits the spurious feature, performance is very high on train and validation, where the shortcut is perfectly predictive, but collapses on the unbiased test set, where the shortcut is uncorrelated or no longer aligned (Won et al., 2 Oct 2025).

5. Observed shortcut learning and generalization failure

The empirical results illustrate the intended failure mode with unusual clarity. For a baseline setting without spurious correlation and with large data, a ResNet-50 attains train accuracy of 0.80, validation accuracy of 0.76, and test accuracy of 0.82 (Won et al., 2 Oct 2025). Under the magnetic field strength split, the same model reaches 0.99 train accuracy and 0.99 validation accuracy, but only 0.52 test accuracy. Under the vertical alignment split, it reaches 0.99 train accuracy and 1.00 validation accuracy, again with 0.52 test accuracy on the unbiased test set (Won et al., 2 Oct 2025).

The weaker spurious variables reported in the paper show a less extreme but still notable pattern. For ethnicity, the reported ResNet-50 accuracies are 0.97 on train, 0.85 on validation, and 0.72 on test. For menopause, they are 0.91 on train, 0.85 on validation, and 0.71 on test (Won et al., 2 Oct 2025). These results indicate that some spurious features induce partial rather than near-total shortcut dependence.

A common misunderstanding in medical AI is that very high internal validation accuracy necessarily reflects clinically relevant representation learning. SpurBreast provides a counterexample by construction. In the magnetic field and orientation settings, near-perfect validation accuracy is compatible with test accuracy close to random guessing on the unbiased set (Won et al., 2 Oct 2025). The paper’s central point is therefore not simply that spurious correlations exist, but that standard validation protocols can reward them.

6. Methodological role and research uses

SpurBreast is positioned as a benchmark for studying clinically relevant and clinically irrelevant features, uncertainty estimation, adversarial robustness, and generalization strategies (Won et al., 2 Oct 2025). Because it contains both biased and unbiased splits, it can be used to compare models or interventions under conditions where shortcut opportunities are either present or removed. The paper explicitly notes strategies such as domain adaptation, reweighting, and invariant risk minimization as relevant classes of methods for mitigation or removal of shortcut feature reliance (Won et al., 2 Oct 2025).

The dataset also has a reproducibility function. The splits are described as controlled and transparent, and the image data are paired with full documentation and accompanying tabular data (Won et al., 2 Oct 2025). Models and datasets are made available through the project repository cited in the paper, which reinforces the benchmark’s role as shared infrastructure rather than as a single-study case analysis (Won et al., 2 Oct 2025).

Relative to prior medical imaging datasets such as CheXpert, MURA, and MIMIC-CXR, SpurBreast is distinguished by being expressly designed to systematically investigate spurious correlations rather than merely containing them incidentally (Won et al., 2 Oct 2025). Relative to synthetic benchmarks, it preserves real clinical MRI while still allowing explicit control over confounding structure (Won et al., 2 Oct 2025). A plausible implication is that SpurBreast is best understood as a dataset for robustness research in medical imaging: it formalizes the gap between apparent performance and clinically meaningful generalization, and does so within a breast MRI setting where image, protocol, and demographic variables can all participate in model behavior.