⚡ electrical engineering

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

The paper introduces LUMINA, a multi-vendor full-field digital mammography benchmark featuring energy harmonization protocols and rich clinical annotations to address data limitations and improve the robustness and localization accuracy of AI models across diverse acquisition systems.

Original authors: Hongyi Pan, Gorkem Durak, Halil Ertugrul Aktas, Andrea M. Bejar, Baver Tutun, Emre Uysal, Ezgi Bulbul, Mehmet Fatih Dogan, Berrin Erok, Berna Akkus Yildirim, Sukru Mehmet Erturk, Ulas Bagci

Published 2026-03-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hongyi Pan, Gorkem Durak, Halil Ertugrul Aktas, Andrea M. Bejar, Baver Tutun, Emre Uysal, Ezgi Bulbul, Mehmet Fatih Dogan, Berrin Erok, Berna Akkus Yildirim, Sukru Mehmet Erturk, Ulas Bagci

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a computer to spot breast cancer in X-ray images (mammograms). You'd think this would be easy: just feed the computer thousands of pictures, and it learns the pattern, right?

But in the real world, it's like trying to teach a student to recognize a specific type of apple by showing them apples from six different grocery stores. One store sells apples that are shiny and red (Vendor A), another sells dull, green ones (Vendor B), and a third sells apples that look like they were taken with a flash (High Energy) while another uses natural light (Low Energy).

If you train your student only on the shiny red apples, they might fail when they see a dull green one, even if it's the same kind of apple. This is exactly the problem doctors and AI researchers face with mammograms. Different machines and settings make the images look totally different, confusing the AI.

This paper introduces LUMINA, a solution to this problem, along with a clever new way to fix the images before the AI sees them.

1. The New Library: LUMINA

Think of previous datasets as small, dusty libraries with only a few books from one publisher. They were too small and too similar to teach a smart AI how to handle the real world.

LUMINA is a massive, diverse library.

The Collection: It contains 1,824 high-quality digital mammograms from 468 real patients.
The Diversity: Unlike old datasets, LUMINA includes images from six different machine manufacturers (like Siemens, GE, Fujifilm) and captures images taken with different energy settings (like different camera flashes).
The Labels: Every single image is tagged with the "truth": did the patient have cancer? What was the risk level? How dense is the breast tissue? This is the "answer key" the AI needs to learn.

2. The Magic Trick: "Energy Harmonization"

Here is the paper's biggest innovation. Before the AI looks at the images, the researchers perform a magic trick called Foreground-Only Harmonization.

The Problem:
Mammograms have a huge black background (the air around the breast) and a bright white/grey foreground (the breast tissue).

Old Method: If you try to "normalize" the image (make the colors look consistent), the computer gets confused by the massive black background. It's like trying to adjust the volume on a radio, but the static noise is so loud it drowns out the music. The computer tries to match the black background, which ruins the details of the breast tissue.

The New Method:
The researchers invented a filter that says, "Ignore the black background. Only look at the breast."

They take an image from a "High Energy" machine (which looks a certain way) and mathematically stretch its colors to match a "Low Energy" reference image, but only for the breast tissue.
The black background stays black. The breast tissue gets "harmonized" to look like it came from the same machine.

The Result:
It's like taking photos from six different cameras and running them through a filter that makes them all look like they were taken with the same high-quality lens, without blurring the important details.

3. The Test Drive: What Did They Find?

The researchers put this new library and the new filter to the test using three different "driving tests" for the AI:

Diagnosis: Is it cancer or not?
Risk Level: How worried should the doctor be? (BI-RADS scores)
Density: How thick is the breast tissue?

The Winners:

Two Views are Better than One: Just like looking at a sculpture from the front and the side gives you a better understanding, the AI performed best when it saw two views of the breast (top-down and side-angle) together.
The Best Models: A model called EfficientNet-B0 was the champion for spotting cancer, and Swin-T was the best at judging tissue density.
The Harmonization Effect: When they used the "Magic Filter" (harmonization), the AI got smarter. It made fewer mistakes, and when they looked at where the AI was looking (using a heat map), it focused much more sharply on the suspicious spots and stopped getting distracted by the edges of the image or the chest wall.

The Big Picture

This paper is a game-changer because it solves the "apples and oranges" problem in medical AI.

Before: AI models were fragile. If you trained them on Machine A, they broke on Machine B.
Now: With LUMINA and the new harmonization trick, we have a blueprint for building AI that works reliably across different hospitals and different machines.

It's like finally giving the AI a universal translator that allows it to understand every dialect of "breast X-ray," ensuring that no matter where a patient goes, the AI can help the doctor make the right call.

1. Problem Statement

Current deep learning models for breast cancer detection face significant challenges in generalization due to the limitations of existing public mammography datasets:

Data Scarcity & Heterogeneity: Existing datasets (e.g., CBIS-DDSM, INbreast) are often limited in size, rely on older screen-film mammography (SFM), or are single-vendor, failing to capture the diversity of modern Full-Field Digital Mammography (FFDM).
Domain Shifts: Variations in acquisition energy (low vs. high energy) and vendor-specific post-processing algorithms create substantial intensity and contrast discrepancies. These "domain shifts" degrade model performance when deployed across different clinical settings.
Lack of Comprehensive Labels: Many datasets lack complete pathology confirmation, BI-RADS risk assessments, and breast density annotations simultaneously, hindering multi-task learning.

2. Methodology

A. The LUMINA Dataset

The authors introduce LUMINA, a curated, multi-vendor FFDM dataset designed to address these gaps.

Scale & Composition: 1,824 images from 468 patients (250 benign, 218 malignant).
Annotations: Pathology-confirmed outcomes, expert-assigned BI-RADS risk categories (0–6), and breast density classifications (A–D).
Diversity: Images are sourced from six different vendors (IMS, Metaltronica, FUJIFILM, Siemens, Carestream, GE) with varying acquisition energies (High and Low).
Resolution: Native resolutions range from 2364×2964 to 4800×6000 pixels with 12–14 bit depth.
Preprocessing: FUJIFILM images (MONOCHROME1) were inverted to MONOCHROME2 to ensure consistent visualization.

B. Foreground-Only Pixel-Space Harmonization

To mitigate vendor and energy-driven appearance shifts without altering lesion morphology, the authors propose a novel foreground-only histogram harmonization protocol:

Concept: Standard histogram matching fails in mammography because the large black background (intensity 0) skews the distribution.
Mechanism:
1. Masking: Only pixels with intensity $>0$ (the breast tissue/foreground) are considered.
2. CDF Mapping: Cumulative Distribution Functions (CDFs) are computed only for the foreground pixels of the source image and a low-energy reference image.
3. Alignment: A mapping function $T(\cdot)$ aligns the source CDF to the reference CDF.
4. Application: The mapping is applied to foreground pixels, while the background remains zero.
Goal: This normalizes intensity distributions across vendors to a unified low-energy style while preserving the contrast and morphology of lesions.

C. Deep Learning Framework

The study benchmarks modern architectures on three clinically relevant tasks:

Diagnosis: Binary classification (Benign vs. Malignant).
BI-RADS Risk Grouping: Binary (Low vs. High risk) and 3-class grouping.
Breast Density: 4-class prediction (A, B, C, D).

Architectures: EfficientNet-B0, ResNet-50, DenseNet-121, and Swin-Transformer (Swin-T).
View Strategy: Comparison of Single-View (CC or MLO) vs. Two-View (CC + MLO concatenated) models. The two-view models utilize a shared-weight backbone to improve parameter efficiency.

3. Key Contributions

LUMINA Dataset: The first multi-vendor, energy-annotated FFDM benchmark with comprehensive pathology, BI-RADS, and density labels, enabling controlled stress-testing for domain shifts.
Harmonization Protocol: A model-agnostic, lightweight pre-processing step (foreground-only CDF matching) that effectively reduces cross-vendor/energy drift without complex generative models.
Unified Multi-Task Benchmark: A rigorous evaluation framework comparing single-view vs. two-view modeling across three distinct clinical tasks.
Insight on Attention: Demonstration that harmonization not only improves metrics but also enhances the spatial focus of model attention (Grad-CAM) on suspicious regions.

4. Experimental Results

Performance Metrics (AUC)

Diagnosis (Benign vs. Malignant):
- Best Model: EfficientNet-B0 (Two-View, 512² input).
- Result: 93.54% AUC.
- Finding: Two-view models consistently outperformed single-view models across all backbones.
BI-RADS Classification:
- Best Model: EfficientNet-B0.
- Result: 92.80% AUC (Binary) and 83.27% AUC (3-class).
Breast Density:
- Best Model: Swin-T.
- Result: 89.43% AUC (at 224² input), outperforming CNNs.

Impact of Harmonization

Metric Improvement: Harmonization consistently improved Accuracy (ACC), AUC, and F1-scores across all backbones and tasks.
- Example: Two-view diagnosis AUC improved from 91.42% (Raw) to 92.99% (Harmonized) for EfficientNet-B0.
Attention Localization: Grad-CAM visualizations showed that harmonized models produced more focal attention maps centered on lesions, whereas raw models often attended to background artifacts or chest walls.
Energy Specificity: Harmonization significantly reduced the performance gap between low-energy and high-energy subsets, with low-energy images showing the most significant gains.

Efficiency

Shared vs. Independent Backbones: Using a shared-weight backbone for two-view models reduced parameters by 48% (from 8.34M to 4.34M) while maintaining comparable or slightly better AUC performance compared to independent backbones.

5. Significance

Clinical Reliability: LUMINA provides a realistic benchmark for developing AI that can generalize across different hospital systems and scanner types, a critical requirement for clinical deployment.
Methodological Advancement: The foreground-only harmonization protocol offers a simple, effective solution to a complex domain adaptation problem, proving that pre-processing can be as impactful as architectural changes.
Multi-View Superiority: The study confirms that combining CC and MLO views is essential for high-performance mammography AI, validating the clinical standard of two-view imaging.
Resource Availability: By releasing the dataset and code, the authors facilitate reproducible research and accelerate the development of robust, deployable breast cancer screening AI.

In conclusion, LUMINA establishes a new standard for mammography AI research by addressing the critical issues of vendor diversity and energy variation through a unique dataset and a novel harmonization protocol, resulting in more accurate and interpretable diagnostic models.