BreastMammo and DenseMammo: Benchmarks for Mammography Domain Generalization

Breast density classification is a critical component of breast cancer risk assessment, yet AI models often struggle to generalize across clinical sites due to vendor-specific acquisition styles. In this work, we introduce two new datasets, BreastMammo and DenseMammo, to facilitate robust multi-view mammography research. We propose a domain generalization framework that utilizes a foreground-only histogram matching protocol to resolve the domain shift issue arising from disparate clinical sources. Internal evaluation using a 5-fold cross-validation protocol demonstrates the efficacy of our approach, with the Swin Transformer backbone achieving a peak AUC of 98.32% for density classification. External evaluation on the TNMammo and LUMINA datasets demonstrates that the proposed approach consistently reduces domain shift, significantly outperforming prominent domain generalization paradigms, including MixStyle and Discrete-Fourier-Transform-based frameworks.

BreastMammo is a two-view diagnostic dataset comprising 894 images from 447 patients. Each patient profile includes paired Craniocaudal (CC) and Mediolateral Oblique (MLO) views of a single breast. The cohort consists of 576 benign and 318 malignant images with pathology-confirmed labels, BI-RADS assessment scores, and breast density categories (ACR categories A–D). DenseMammo is a large-scale four-view screening dataset for breast density classification. It contains 2,480 images from 620 patients. Adhering to standard clinical protocols, each patient provided a full screening set consisting of four views: Left CC, Left MLO, Right CC, and Right MLO. All images are provided with expert-annotated breast density labels.

Our domain generalization approach builds upon the histogram-based data harmonization framework introduced by LUMINA, which demonstrated the efficacy of foreground-only pixel-space alignment in mitigating vendor-induced intensity drift. While LUMINA utilizes histogram to map images to a specific low-energy reference target for static harmonization, we extend this principle to a target-free domain generalization framework that dynamically bridges domain gaps by decoupling breast tissue style from its morphological content during training. Following traditional histogram matching protocols, given a source image Is from domain Ds and a reference image Ir from domain Dr, an intensity mapping function T is established by aligning their cumulative distribution functions (CDFs). Crucially, because mammograms contain large, non-clinical black background regions that bias statistical counts and degrade matching quality, these uninformative regions are explicitly excluded by strictly evaluating pixels with an intensity value, p, greater than zero (p > 0). The histograms and their normalized CDFs (C̄s and C̄r) are subsequently computed solely over these isolated breast tissue foreground masks, mapping the source pixel intensities such that C̄s(p) ≈ C̄r(T(p)), where T(·) denotes the intensity mapping function. Moreover, we follow Liu et al. to introduce a hyperparameter α ∈ [0, 1] to control the intensity of the style shift for the final synthetic image Io by linearly blending the matched result with the original source characteristics: Io(x, y) = α · T(Is(x, y)) + (1 − α) · Is(x, y), where (x, y) denote the spatial coordinates. By varying α, we generate a series of synthetic samples that balance the original source characteristics with the target domain’s intensity profile. During training, our pairing mechanism operates under two strict clinical constraints to ensure data consistency. First, for each training instance originating from a source domain, a reference image is dynamically and randomly sampled from the alternative target domain’s training pool to maximize cross-institutional style exposure. Second, because clinical mammography relies on multi-view examinations, we enforce inter-view style synchronization that the identical reference style target is utilized across all corresponding views (CC and MLO) for a given patient case. For each raw patient case, we generate four synthetic variations spanning progressive style strengths (α ∈ 0.25, 0.5, 0.75, 1) to establish a continuous training manifold.

The aligned images are processed through a Swin-T backbone. For multi-view analysis, the pipeline implements shared weights across CC and MLO views. To ensure spatial consistency across bilateral images, we followed MammoClean to horizontally flip all right-side images before feature extraction. The hierarchical shifted-window attention of the Swin-T allows the model to capture fine pathological textures while maintaining a global context of the breast structure. Following feature extraction, the latent representations from each view are concatenated into a single feature vector. This unified representation is passed through a classification head consisting of linear layers to produce predictions for our clinical tasks: (1) density assessment, where models categorize breasts from BreastMammo and DenseMammo into one ACR density category (A, B, C, or D); and (2) pathology diagnosis, where models distinguish BreastMammo’s breasts between benign and malignant cases based on the fused two-view features.

For internal evaluation, we benchmarked several state-of-the-art backbones, including EfficientNet-B0, DenseNet-121, ResNet-50, and Swin-T. All backbones were pretrained on the ImageNet-1K dataset under 2242 resolution. Five-fold cross validation was adopted to establish a standard baseline benchmark. The internal evaluation was conducted using standard training protocols without any domain generalization techniques. The internal evaluation results demonstrate that the Swin-T architecture generally provides the most robust baseline. In the density classification task, Swin-T consistently outperformed CNN-based methods in AUC (96.53% on BreastMammo and 98.32% on DenseMammo). Notably, utilizing a larger input resolution (5122) did not yield significant performance gains. This marginal improvement is likely attributed to a resolution mismatch, as the foundational backbones were initialized with weights pretrained on the ImageNet-1K dataset at a native 2242 resolution.

External evaluation was conducted to evaluate our proposed histogram-based domain generalization in mitigating multi-center domain shifts using unseen datasets TNMammo and LUMINA. Swin-T was selected as the backbone due to its superior internal performance. MixStyle (feature-level augmentation) and DFT-based domain generalization (low-frequency amplitude spectrum components swapping between source and reference samples) were implemented for comparison. Crucially, because vanilla Fourier swapping introduces severe uninformative noise across the background, we implemented an optimized version that enforces a post-processing spatial mask to reset background pixels to zero where the original source image is zero. Experimental evaluation on unseen datasets demonstrates the robust generalization achieved by our approach. Our protocol successfully mitigates domain shifts by aligning pixel-space distributions across diverse vendor acquisition styles, allowing the multi-view Swin-T model to retain high diagnostic fidelity on completely novel institutional data. Specifically, our method consistently achieved top-tier performance across both external datasets. It improved the AUC from 83.46% to 86.38% on the TNMammo dataset and from 81.72% to 86.13% on the LUMINA dataset, with corresponding consistent gains in ACC and F1-score. On the contrary, alternative domain generalization methods showed limited or inconsistent efficacy.

The paper discusses why data expansion surpasses feature-space manipulation: While MixStyle introduces regularization by implicitly blending feature-map statistics within individual mini-batches, this abstract manipulation lacks explicit anatomical grounding and can inadvertently corrupt the delicate, localized texture patterns critical for distinguishing tissue densities. In contrast, our foreground-only histogram matching protocol demonstrates that explicitly expanding the training manifold via high-fidelity, pixel-space synthetic images is far more efficient and structurally reliable. By directly generating a continuous spectrum of tissue-specific intensity profiles, we expose the network to diverse vendor characteristics while strictly preserving the underlying anatomical structures. This data expansion allows models to naturally learn robust, domain-invariant representations from intact clinical morphology, resulting in superior feature scaling when deployed to entirely unseen external institutions.

The paper also discusses why foreground histogram matching outperforms DFT on mammograms: The Fourier-based domain generalization takes the massive and uninformative black background into the low-frequency style calculation. By contrast, our model-agnostic histogram matching protocol operates strictly on dynamically isolated foreground pixels. By completely ignoring the background from the outset, it ensures that intensity alignment is purely driven by actual breast tissue density distributions. This continuous blending effectively harmonizes vendor-specific contrast variances without introducing frequency leakage or artificial edge discontinuities, directly leading to the superior generalization performance.

In conclusion, this work presented a robust, anatomically grounded framework for multi-view mammography analysis designed to overcome the pervasive challenges of vendor-induced domain shift. To support the research community in developing reliable clinical models, we introduced and benchmarked two distinct multi-view FFDM datasets BreastMammo and DenseMammo. By integrating a novel foreground-only histogram matching protocol with a two-view Swin-T architecture, we successfully reduced the mammography domain shift on the external datasets. Our external validation on the unseen TNMammo and LUMINA datasets demonstrates that our localized, pixel-space alignment strategy consistently improves model generalization, significantly outperforming traditional frequency-space interpolation and feature-level statistical blending paradigms. These results underscore the critical importance of tissue-specific intensity alignment for building scalable, site-agnostic deep learning frameworks capable of managing the masking effect in diverse clinical screening environments.

Improvements to AI Systems:

1. Implement foreground-only histogram matching as a preprocessing layer – Replace or augment feature-space domain adaptation (e.g., MixStyle) with pixel-space intensity alignment that strictly operates on non-zero foreground pixels (breast tissue only), ignoring black background regions. This reduces vendor-induced intensity drift without corrupting localized texture patterns.

2. Use dynamic cross-domain style augmentation with controlled blending – During training, randomly sample reference images from alternative domains and generate synthetic samples by linearly blending matched histograms with the original image using α ∈ 0.25, 0.5, 0.75, 1. This creates a continuous training manifold that exposes the model to diverse vendor-specific intensity profiles while preserving anatomical structure.

3. Enforce multi-view style synchronization – For multi-view inputs (CC and MLO), apply the identical reference style target across all views of the same patient. This ensures inter-view consistency, which is critical for clinical tasks that rely on paired views.

4. Adopt a Swin Transformer backbone with shared weights across views – Use hierarchical shifted-window attention to capture both fine pathological textures and global breast structure. Concatenate latent features from each view before classification, enabling joint reasoning across CC and MLO perspectives.

5. Add horizontal flipping for right-side images – Standardize spatial orientation across bilateral images (flip right-side images) to ensure consistent feature extraction, as done in MammoClean.

6. Optimize Fourier-based methods by masking background – If using DFT-based domain generalization, enforce a post-processing spatial mask to reset background pixels to zero where the source image is zero, preventing uninformative noise from low-frequency background regions.

What the Improved AI System Can Do:

• Generalize across unseen clinical sites – Achieve higher AUC on external datasets (e.g., improve from 83.46% to 86.38% on TNMammo and from 81.72% to 86.13% on LUMINA) by reducing vendor-induced domain shift.

• Perform robust breast density classification – Accurately categorize breasts into ACR categories (A–D) with up to 98.32% AUC on internal data, maintaining high fidelity on novel institutional data.

• Conduct reliable pathology diagnosis – Distinguish benign vs. malignant cases from two-view mammograms using fused features, with improved sensitivity to subtle tissue differences.

• Handle multi-view clinical protocols – Process both two-view (CC/MLO) and four-view (bilateral CC/MLO) screening sets while preserving inter-view consistency.

• Resist background noise artifacts – Avoid performance degradation caused by large non-clinical black regions in mammograms, which commonly bias statistical calculations and degrade style transfer quality.

• Scale to diverse acquisition styles – Maintain diagnostic performance across different vendors, imaging protocols, and contrast settings without requiring site-specific fine-tuning.

Breast density classification is a critical component of breast cancer risk assessment, yet AI models often struggle to generalize across clinical sites due to vendor-specific acquisition styles. In this work, we introduce two new datasets, BreastMammo and DenseMammo, to facilitate robust multi-view mammography research. We propose a domain generalization framework that utilizes a foreground-only histogram matching protocol to resolve the domain shift issue arising from disparate clinical sources. Internal evaluation using a 5-fold cross-validation protocol demonstrates the efficacy of our approach, with the Swin Transformer backbone achieving a peak AUC of 98.32% for density classification. External evaluation on the TNMammo and LUMINA datasets demonstrates that the proposed approach consistently reduces domain shift, significantly outperforming prominent domain generalization paradigms, including MixStyle and Discrete-Fourier-Transform-based frameworks.