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

Hongyi Pan; Gorkem Durak; Halil Ertugrul Aktas; Andrea M. Bejar; Baver Tutun; Emre Uysal; Ezgi Bulbul; Mehmet Fatih Dogan

eess.IVMar 15, 2026

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

Authors: Hongyi Pan, Gorkem Durak, Halil Ertugrul Aktas, Andrea M. Bejar, Baver Tutun, Emre Uysal, Ezgi Bulbul, Mehmet Fatih Dogan, +4 more

Organizations: Department of Radiology, Northwestern University, Chicago, IL, USA · Department of Radiation Oncology, University of Health Sciences Prof. Dr. Cemil Tascioglu City Hospital, Istanbul, Turkey · Department of Radiology, Istanbul University, Istanbul, Turkey

Abstract

Publicly available full-field digital mammography (FFDM) datasets remain limited in size, clinical annotations, and vendor diversity, hindering the development of robust models. We introduce LUMINA, a curated, multi-vendor FFDM dataset that explicitly encodes acquisition energy and vendor metadata to capture clinically relevant appearance variations often overlooked in existing benchmarks. This dataset contains 1824 images from 468 patients (960 benign, 864 malignant), with pathology-confirmed labels, BI-RADS assessments, and breast-density annotations. LUMINA spans six acquisition systems and includes both high- and low-energy imaging styles, enabling systematic analysis of vendor- and energy-induced domain shifts. To address these variations, we propose a foreground-only pixel-space alignment method (''energy harmonization'') that maps images to a low-energy reference while preserving lesion morphology. We benchmark CNN and transformer models on three clinically relevant tasks: diagnosis (benign vs. malignant), BI-RADS classification, and density estimation. Two-view models consistently outperform single-view models. EfficientNet-B0 achieves an AUC of 93.54% for diagnosis, while Swin-T achieves the best macro-AUC of 89.43% for density prediction. Harmonization improves performance across architectures and produces more localized Grad-CAM responses. Overall, LUMINA provides (1) a vendor-diverse benchmark and (2) a model-agnostic harmonization framework for reliable and deployable mammography AI.

Figures & tables

Figure 1: Representative benign and malignant mammograms.

				Pathology	BI-RADS	Breast	Acquisition
Dataset	Origin	Patients	Images	Labels	Assessment	Density	Technique
MIAS [ 30 ]	UK	161	322	Yes	No	No	SFM
DDSM [ 4 ]	USA	2,620	10,480	Yes	Yes	Yes	SFM
CBIS-DDSM [ 14 ]	USA	2,620	10,239	Yes	Yes	Yes	SFM
INbreast [ 19 ]	Portugal	115	410	No	Yes	Yes	FFDM
VinDR-Mammo [ 22 ]	Vietnam	5,000	20,000	No	Yes	Yes	FFDM

Table 1: Comparison of publicly available mammography datasets.

Panel of Figure 2 — Figure 2: Age, BI-RADS, and breast distribution.

Vendor	# Patients	# Images	Energy
IMS	341	1326	High
Metaltronica	91	354	Low
FUJIFILM Corporation	29	116	Low
SIEMENS	4	16	Low
Carestream Health	1	4	Low
GE MEDICAL SYSTEMS	2	8	High

Table 2: Vendor distribution.

Figure 3: LUMINA pipeline. The EfficientNet-B0 backbone can be replaced by other backbones (ResNet-50, DenseNet-121, and Swin-T). Two‑view shared‑backbone reached accuracy comparable to independent‑backbone with 48% less parameters (Table 7 ) and outperformed four‑view variants (Table 4 ).

View	Benign	Malignant	Total
Single	592 (592)	344 (344)	936 (936)
Two	296 (592)	172 (344)	468 (936)

Table 3: Breast cancer diagnosis task data distribution. Numbers outside and inside brackets are cases and images, respectively.

Figure 4: Background influence. Standard histogram matching is degraded by large black background regions, whereas foreground-only histogram matching remains unaffected.

View	Model	Input	Params	Flops	ACC(%)	AUC(%)	Precision(%)	Recall(%)	F1(%)	Specificity(%)
Single	ResNet-50	$224^{2}$	23.51M	4.13G	79.49 $\pm$ 5.99	87.42 $\pm$ 4.21	81.29 $\pm$ 10.90	61.97 $\pm$ 18.17	67.52 $\pm$ 13.09	89.68 $\pm$ 9.63
	DenseNet-121	$224^{2}$	6.95M	2.90G	79.49 $\pm$ 3.37	87.14 $\pm$ 3.08	80.45 $\pm$ 9.45	62.55 $\pm$ 17.56	67.79 $\pm$ 9.57	89.38 $\pm$ 7.52
	EfficientNet-B0	$224^{2}$	4.01M	0.41G	83.12 $\pm$ 4.13	90.66 $\pm$ 4.21	79.78 $\pm$ 6.31	73.89 $\pm$ 15.77	75.50 $\pm$ 7.95	88.51 $\pm$ 5.00
	Swin-T	$224^{2}$	27.52M	3.13G	81.84 $\pm$ 5.92	88.62 $\pm$ 5.18	78.27 $\pm$ 8.63	71.83 $\pm$ 12.10	74.16 $\pm$ 8.36	87.66 $\pm$ 7.45
	ResNet-50	$512^{2}$	23.51M	21.58G	77.68 $\pm$ 7.18	86.69 $\pm$ 4.37	82.79 $\pm$ 14.09	59.36 $\pm$ 28.73	61.41 $\pm$ 22.78	88.38 $\pm$ 12.29
	DenseNet-121	$512^{2}$	6.95M	15.14G	79.38 $\pm$ 5.55	90.52 $\pm$ 3.61	81.32 $\pm$ 16.39	68.09 $\pm$ 21.96	69.22 $\pm$ 11.98	85.99 $\pm$ 14.42

Table 4: Mammogram cancer diagnosis benchmark. The best and second-best results are highlighted in green and yellow, respectively.

		Two-Class (BI-RADS 1/2/3 VS 4/5/6)			Three-Class (BI-RADS 1/2 VS 3/4 VS 5/6)
Model	Input	ACC(%)	AUC(%)	F1(%)	ACC(%)	AUC(%)	F1(%)
ResNet-50	$224^{2}$	71.98 $\pm$ 12.34	88.23 $\pm$ 3.52	62.45 $\pm$ 10.33	69.02 $\pm$ 5.14	79.58 $\pm$ 3.00	55.36 $\pm$ 1.97
DenseNet-121	$224^{2}$	77.55 $\pm$ 3.67	87.89 $\pm$ 3.52	58.98 $\pm$ 14.70	68.15 $\pm$ 3.12	80.90 $\pm$ 3.74	48.63 $\pm$ 3.49
EfficientNet-B0	$224^{2}$	83.56 $\pm$ 4.76	92.80 $\pm$ 4.13	74.18 $\pm$ 9.32	71.78 $\pm$ 5.03	83.27 $\pm$ 4.91	55.15 $\pm$ 4.65
Swin-T	$224^{2}$	79.29 $\pm$ 9.01	90.80 $\pm$ 4.61	60.82 $\pm$ 31.02	70.51 $\pm$ 4.45	81.69 $\pm$ 5.55	49.55 $\pm$ 6.85
ResNet-50	$512^{2}$	75.22 $\pm$ 4.46	89.08 $\pm$ 5.42	64.78 $\pm$ 9.96	66.66 $\pm$ 3.10	78.92 $\pm$ 4.91	50.71 $\pm$ 3.73

Table 5: BI-RADS classification benchmark.

Model	Input	ACC(%)	AUC(%)	F1(%)
ResNet-50	$224^{2}$	64.74 $\pm$ 3.33	85.66 $\pm$ 1.62	56.94 $\pm$ 6.43
DenseNet-121	$224^{2}$	66.29 $\pm$ 4.20	86.39 $\pm$ 2.92	56.70 $\pm$ 2.97
EfficientNet-B0	$224^{2}$	59.43 $\pm$ 5.23	84.72 $\pm$ 3.54	48.25 $\pm$ 5.68
Swin-T	$224^{2}$	70.35 $\pm$ 4.60	89.43 $\pm$ 2.22	65.28 $\pm$ 6.48
ResNet-50	$512^{2}$	65.37 $\pm$ 4.07	87.06 $\pm$ 2.10	58.29 $\pm$ 7.57
DenseNet-121	$512^{2}$	66.31 $\pm$ 5.31	86.60 $\pm$ 3.76	61.42 $\pm$ 7.62

Table 6: Density classification benchmark.

Panel of Figure 5 — Figure 5: t-SNE visualization for diagnosis task. Higher resolution increases benign/malignant separation in the latent space.

Backbone	Input	Params	ACC(%)	AUC(%)	F1(%)
Indepen-	$224^{2}$	8.34M	84.82 $\pm$ 3.00	93.54 $\pm$ 2.38	77.41 $\pm$ 6.13
dent	$512^{2}$	8.34M	88.46 $\pm$ 4.91	92.71 $\pm$ 4.63	82.19 $\pm$ 8.61
Shared	$224^{2}$	4.34M	86.11 $\pm$ 2.47	92.99 $\pm$ 3.04	79.39 $\pm$ 4.18
	$512^{2}$	4.34M	85.04 $\pm$ 5.92	93.54 $\pm$ 3.88	75.75 $\pm$ 12.85

Table 7: Shared vs. independent two-view EfficientNet-B0 backbones. Flops are 0.82G for

224^{2}

and 4.27G under

512^{2}

Panel of Figure 6 — Figure 6: Harmonization improves ACC, AUC, and F1-Score across tasks. Bars show mean $\pm$ std over 5 folds, and curves show the mean ROC. How to read: In (a) and (b), darker bars (harmonized) consistently exceed lighter bars (raw). In (c) and (d), areas under the blue and orange (harmonized) curves are respectively larger than areas under the green and red (raw) curves. Takeaway: Pixel-space CDF alignment helps regardless of backbone or input size, supporting its use as a robust pre-processing step.

Panel of Figure 7 — Figure 7: Attention becomes more focal after harmonization. Shown are raw (left two columns) vs. harmonized (right two columns) inputs with Grad-CAM overlays for two malignant cases. Observation: Harmonization reduces diffuse activations and concentrates attention on lesion-bearing regions. Implication: Harmonization not only improves metrics (Fig. 6 ) but may also enhance clinical interpretability by focusing on suspicious tissue.

Figure 8: Energy-specific AUC. Histogram harmonization improved AUC, particularly for low-energy images.

Panel of Figure 9 — Figure 9: ROC curves of the EfficientNet-B0 model for breast cancer diagnosis.

Panel of Figure 10 — Figure 10: ROC curves of the EfficientNet-B0 model for two-class BI-RADS classification.

Family	Examples	Labels/Supervision	Architecture/Objective	Limitations vs. LUMINA
Film (SFM) datasets	MIAS [ 30 ] , DDSM [ 4 ] , CBIS-DDSM [ 14 ]	Pathology (yes), BI-RADS (varies)	Early CNNs / CAD; single-task classification	Film scans, lower resolution; limited modern FFDM relevance.
Digital (FFDM) datasets	INbreast [ 19 ] , VinDR-Mammo [ 22 ] , RSNA [ 6 ] , CMMD [ 5 ] , KAU-BCMD [ 1 ]	Pathology or BI-RADS; density varies	CNNs/ViTs; single- or dual-task	Often single vendor/system; partial labels; limited multi-task scope.
Harmonization (feature space)	ComBat [ 21 ]	Batch-effect correction on features	Empirical Bayes (location/scale) on radiomics/latent features	Requires feature extraction; not pixel-space; modality-agnostic assumptions.
Harmonization (federated learning)	HarmoFL [ 12 ]	Federated frequency‑domain drift normalization	Joint optimization across clients	Heavy infra; task/model coupling; not a simple preproc.
LUMINA (Ours)	Multi-vendor FFDM (6 systems); energy metadata	Pathology + BI-RADS + density	Foreground-only pixel-space CDF matching + unified 3‑task benchmark; single-/two-/four‑view baselines	Vendor/energy diversity; three tasks; consistent harmonization gains; improved attention focus.

Table 8: Positioning. LUMINA provides pathology, BI-RADS, and density on multi-vendor FFDM, plus a simple, effective pixel‑space harmonization baseline.

Backbone	Pre-trained	Learning Rate	Weight Decay	Epochs	Learning Rate scheduler	Batch	Augmentation
ResNet-50	ImageNet-1K	$1\!\times\!10^{-3}$	$1\!\times\!10^{-5}$	100	$\times 0.1$ every 30	32	flip, resize
DenseNet-121	ImageNet-1K	$1\!\times\!10^{-3}$	$1\!\times\!10^{-5}$	100	$\times 0.1$ every 30	32	flip, resize
EfficientNet-B0	ImageNet-1K	$1\!\times\!10^{-3}$	$1\!\times\!10^{-5}$	100	$\times 0.1$ every 30	32	flip, resize
Swin-T	ImageNet-1K	$1\!\times\!10^{-5}$	$1\!\times\!10^{-5}$	100	$\times 0.1$ every 30	32	flip, resize

Table 9: Hyperparameter parity across backbones. Shared schedule and augmentations ensure a fair comparison (see Sec. 6.1 ).

Stage	Operation (applied to all models equally)
DICOM handling	Convert MONOCHROME1 to MONOCHROME2; remove text burn-ins.
Foreground mask	Define $M=\{(x,y)\mid\mathbf{I}(x,y)>0\}$ ; exclude background from histograms.
Harmonization	Foreground CDF matching to low-energy reference (Eqs. 1–8); training-free, model-agnostic.
Resize/replicate	Resize to $224^{2}$ or $512^{2}$ ; replicate grayscale to 3 channels for ImageNet backbones (Sec. 6.1 ).
Augment	Random horizontal flip (all backbones, all tasks).
Normalization	As in backbone defaults; no per-model special tuning beyond LR noted in Sec. 6.1 .

Table 10: Shared pre-processing. Ensures identical inputs across backbones; isolates model differences.

Explore similar work

Aug 10, 2026

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.

Hongyi Pan, Gorkem Durak, Halil Ertugrul Aktas +18

Northwestern University · University of Wisconsin-Madison · Istanbul University +7

Domain Generalization Medical Image Classification Mammography Medical Image Benchmarks

Ask PDF

Jul 11, 2026

Benchmarking the Robustness of Foundation Models for Mammography under Domain Shift

Foundation models are increasingly used as image feature extractors for mammography, but their robustness under external domain shift remains unclear. We benchmark 15 foundation-model backbones across breast density, BI-RADS severity, and cancer status using a unified frozen-backbone linear-probe protocol, training on 3 source datasets and evaluating on 12 task-compatible out-of-distribution (OOD) datasets after label harmonization. Mammography-specific vision-language models (Mammo-FM and MaMA) provide the strongest mean OOD performance, but robustness is not explained by mammography exposure alone. DINOv3 remains a competitive vision-only baseline, and mammography-adapted pretraining does not consistently improve generalization. Dataset-level analysis further shows that even leading models show heterogeneous performance across datasets. Feature-space inspection reveals that useful representations can preserve clinical signal while retaining dataset and acquisition structure. These findings highlight dataset-level OOD evaluation as a central criterion for assessing mammography representations. Our code is publicly available: https://github.com/biomedia-mira/mammo-ood.

Giang Nguyen, Raghav Mehta, Emma A. M. Stanley +4

College of Engineering and Computer Science, VinUniversity, Hanoi, Vietnam · Imperial College London, London, UK · Radiology Department, Vietnam National Cancer Hospital, Hanoi, Vietnam +2

Distribution Shift Robustness Medical Imaging Foundation Models OOD Generalization Medical Image Classification Mammography Medical Image Benchmarks

Ask PDF

Sep 13, 2026

From Density to Biopsy Decisions and Malignancy Prediction: A Benchmark Study of Multimodal Large Language Models Against Radiologists in Digital and Contrast-Enhanced Mammography

Purpose: To compare four multimodal large language models (MLLMs) with radiologists of varying expertise in breast density assessment, BI-RADS assessment, biopsy candidacy determination, and continuous malignancy probability estimation using digital mammography (DM) and contrast-enhanced mammography (CEM). Methods: This study included 179 women with paired DM/CEM examinations and reference standards. Four MLLMs (ChatGPT-5.2, Gemini-3.1 Pro, Sonnet-4.6, Muse Spark) interpreted images with and without masks; three radiologists interpreted non-masked images. Results: For binary density classification on DM, radiologist accuracies ranged from 55.81% to 78.60%, exceeding most MLLM values (62.33%-71.63%), while masks added limited benefit. Five-category BI-RADS accuracies were higher for radiologists on DM (56.74-67.44%) and CEM (62.33-82.79%) compared with MLLMs (DM 31.16-45.12%; CEM 40.00-55.81%). Binary biopsy-candidacy accuracies were likewise higher for radiologists (DM 85.12-89.77%; CEM 86.98-92.09%) than for MLLMs (DM 61.39-75.35%; CEM 69.30-82.79%), although CEM improved performance across all readers. Lesion masks substantially improved MLLM continuous malignancy-probability accuracies from 64.65%-71.63% to 72.56-78.60% on DM and from 67.91%-77.21% to 72.56%-81.86% on CEM, approaching radiologist ranges (DM 63.72-82.79%; CEM 81.86-88.84%). The corresponding AUCs for the top masked models overlapped those of the human readers. Overall, Muse Spark, followed by Sonnet-4.6, demonstrated the strongest performance among the MLLMs across domains. Conclusion: Radiologists generally outperformed MLLMs in categorical tasks, while selected masked models approached human performance for continuous malignancy probability estimation, suggesting a potential adjunctive role.

Ali Abbasian Ardakani, Afshin Mohammadi, Taha Yusuf Kuzan +11

Department of Medicine, Faculty of Medicine and Dentistry, Danube Private University, Krems an der Donau, Austria · Department of Radiology, Faculty of Medicine, Urmia University of Medical Science, Urmia, Iran · Department of Radiology, Yeditepe University, Istanbul, Türkiye +10

Medical VLMs Mammography Breast Cancer Multimodal Model Evaluation Clinical Decision Support

Ask PDF