LUMINA: A Multi-Vendor Mammography Benchmark with Energy Harmonization Protocol

Conference: CVPR 2026 arXiv: 2603.14644 Code: Available Area: Medical Imaging Keywords: mammography, multi-vendor dataset, energy harmonization, histogram matching, benchmark

TL;DR

This paper introduces LUMINA, a multi-vendor full-field digital mammography (FFDM) dataset comprising 468 patients and 1,824 images, accompanied by a foreground-pixel histogram matching protocol for energy harmonization. The benchmark systematically evaluates CNN and Transformer models across three clinical tasks: diagnosis, BI-RADS classification, and breast density prediction.

Background & Motivation

• Background: Existing public mammography datasets (e.g., CBIS-DDSM, INbreast) suffer from notable deficiencies in scale, clinical annotation completeness, and vendor diversity. CBIS-DDSM is derived from legacy screen-film mammography (SFM) scans, while INbreast contains only 115 patients.
• Limitations of Prior Work: Multi-vendor acquisition systems differ in energy settings (high/low energy) and vendor-specific processing pipelines, resulting in substantial domain shift in image appearance and intensity distributions. Consequently, models generalize poorly across vendors.
• Goal: (1) Construct an FFDM benchmark dataset that emphasizes vendor diversity and energy metadata; (2) propose a model-agnostic foreground histogram harmonization method to eliminate vendor/energy-induced domain shift.

Method

Overall Architecture

The LUMINA workflow consists of three stages: (1) Data collection and curation — 1,824 FFDM images from six vendors, with pathology-confirmed malignancy labels, BI-RADS scores, and breast density annotations; (2) Foreground histogram harmonization (Energy Harmonization) — aligning all images to a low-energy reference distribution; (3) Multi-task benchmark evaluation — comparing CNNs (ResNet-50, DenseNet-121, EfficientNet-B0) and a Transformer (Swin-T) across the three clinical tasks.

Key Designs

1. Multi-Vendor Dataset Construction: Data were collected from six vendors — IMS, Metaltronica, FUJIFILM, Siemens, Carestream, and GE — covering 468 patients (250 benign, 218 malignant) in 12–14 bit DICOM format. Annotations include pathology-confirmed outcomes, BI-RADS grades 0–6, and breast density categories A–D. Images in FUJIFILM's MONOCHROME1 format are uniformly converted to MONOCHROME2.

2. Foreground-Only CDF Matching: The core idea is to exclude background pixels (intensity = 0) and perform CDF matching exclusively on the foreground breast region. Specifically, a foreground mask is defined as \(M_s = \{(x,y) \mid \mathbf{I}_s(x,y) > 0\}\); foreground histograms \(H_s(k)\) and \(H_r(k)\) are computed for the source and reference images respectively, normalized into CDFs \(\bar{C}_s(p)\) and \(\bar{C}_r(q)\), and intensity transformation is achieved via the mapping \(\mathcal{T}(p) = \arg\min_q |\bar{C}_s(p) - \bar{C}_r(q)|\). The reference histogram is drawn from the low-energy FFDM subset using 12-bit bins to preserve fine-grained detail. Design Motivation: Standard histogram matching is severely distorted by large areas of black background pixels; the foreground mask effectively mitigates this problem.

3. Dual-View Shared-Backbone Network: CC (craniocaudal) and MLO (mediolateral oblique) views are processed through a weight-sharing backbone independently; the resulting features are concatenated and passed to a fully connected classifier. Compared to independent backbones, weight sharing reduces parameters by 48% (4.34M vs. 8.34M) while achieving comparable or superior performance.

Loss & Training

• Standard cross-entropy classification loss
• AdamW optimizer: \(\text{lr}=1 \times 10^{-3}\) for CNNs; \(\text{lr}=1 \times 10^{-5}\) for Swin-T
• 100 epochs with learning rate decay by 0.1 every 30 epochs; weight decay \(1 \times 10^{-5}\)
• 5-fold cross-validation; best model selected by validation AUC
• Data augmentation limited to horizontal flipping and resizing; grayscale images replicated to three channels
• PyTorch with CUDA determinism flags for reproducibility
• Training environment: 8 × NVIDIA A6000 GPUs

Key Experimental Results

Main Results

Dataset / Task	Metric	Ours (Best)	Prev. SOTA	Notes
Diagnosis (Two-view, 512²)	AUC	93.54% (EfficientNet-B0)	—	Best overall: dual-view + high resolution
Diagnosis (Single-view, 512²)	AUC	92.13% (EfficientNet-B0)	—	Second best among single-view configs
BI-RADS Binary (224²)	AUC	92.80% (EfficientNet-B0)	—	Low/high risk classification
BI-RADS Ternary (224²)	AUC	83.27% (EfficientNet-B0)	—	Low/intermediate/high risk
Density Prediction (224²)	Macro-AUC	89.43% (Swin-T)	—	Transformer better suited for density

Ablation Study

Configuration	Key Metric (AUC)	Notes
Shared backbone EfficientNet-B0 (224²)	92.99%	4.34M parameters
Independent backbone EfficientNet-B0 (224²)	93.54%	8.34M parameters; double the params, marginal gain
Raw images (no harmonization)	Baseline	Lower AUC across all tasks
Foreground histogram harmonization	+Gain	Consistent improvements in ACC/AUC/F1; Grad-CAM more focused

Key Findings

• Dual-view models consistently outperform single-view counterparts, confirming the complementary value of CC and MLO views.
• EfficientNet-B0 achieves the best performance on diagnosis and BI-RADS tasks with only ~4M parameters; Swin-T excels at density prediction.
• Higher input resolution (512²) generally improves performance, though 224² remains competitive with substantially lower computational cost.
• Histogram harmonization not only improves quantitative metrics but also sharpens Grad-CAM attention, directing model focus toward lesion regions.
• Low-energy images benefit most from harmonization, as high-energy images dominate the dataset distribution.

Highlights & Insights

• Practical value of foreground masking: A simple yet effective idea — performing histogram matching after excluding background pixels. Though seemingly straightforward, this design is critical in mammography, where FFDM images contain large areas of black background.
• Model-agnostic preprocessing: The harmonization protocol can be applied as a lightweight preprocessing step to any backbone, making it straightforward to adopt in practice.
• Dataset systematicity: The combination of complete annotations (pathology + BI-RADS + density) with vendor/energy metadata is unique among existing datasets.
• Clinical insight: EfficientNet-B0 wins on diagnostic tasks with the fewest parameters, while Swin-T's global attention mechanism makes it better suited for density prediction — revealing a meaningful relationship between task type and model selection.

Limitations & Future Work

• The dataset scale remains modest (468 patients) compared to large-scale resources such as EMBED (~500K images).
• Data originate from a single institution in Turkey, limiting patient population diversity.
• The reference distribution for harmonization is a representative low-energy FFDM subset; no adaptive reference selection mechanism is explored.
• More advanced domain adaptation methods (e.g., adversarial training, frequency-domain alignment) are not investigated.
• No direct experimental comparison with established cross-vendor harmonization methods (e.g., ComBat, HarmoFL) is provided.
• The four-view model underperforms the dual-view model, likely due to overfitting caused by excessive parameters on a small dataset.

Related Work & Insights

• ComBat corrects batch effects via empirical Bayes in feature space rather than pixel space.
• HarmoFL reduces cross-site variation through frequency-domain amplitude normalization in federated learning settings.
• LUMINA complements datasets such as VinDr-Mammo (5,000 patients, single vendor, Vietnam) and RSNA (1,970 patients) — smaller in scale but offering greater vendor diversity.
• The pixel-space approach proposed here is more interpretable and requires no training, making it complementary to feature-space methods.
• Insight: For multi-center medical imaging research, lightweight pixel-space preprocessing may be more practical than complex domain adaptation pipelines.
• Combining LUMINA with MIL-PF (also CVPR 2026) is a promising direction: applying LUMINA's harmonization as preprocessing followed by frozen encoder + MIL classification.

Rating

• Novelty: ⭐⭐⭐ — Solid dataset contribution, though the proposed method (foreground histogram matching) is technically straightforward.
• Experimental Thoroughness: ⭐⭐⭐⭐ — Covers three tasks, multiple models, multiple resolutions, ablations, visualizations, and energy-level analysis.
• Writing Quality: ⭐⭐⭐⭐ — Rich tables and figures, transparent experimental setup, and a persuasive dataset comparison table.
• Value: ⭐⭐⭐⭐ — The multi-vendor benchmark makes a direct contribution to the community and is publicly released on OSF, Kaggle, and GitHub.

Related Papers

• [CVPR 2026] MIL-PF: Multiple Instance Learning on Precomputed Features for Mammography Classification
• [CVPR 2026] A protocol for evaluating robustness to H&E staining variation in computational pathology models
• [CVPR 2026] MedGEN-Bench: Contextually Entangled Benchmark for Open-Ended Multimodal Medical Generation
• [NeurIPS 2025] Pancakes: Consistent Multi-Protocol Image Segmentation Across Biomedical Domains
• [NeurIPS 2025] Energy Matching: Unifying Flow Matching and Energy-Based Models for Generative Modeling