DBT-2026, a de-identified publicly available dataset of digital breast tomosynthesis exams with ground truth biopsies

This paper introduces DBT-2026, a freely available, de-identified dataset comprising 558 digital breast tomosynthesis exams with expert annotations and clinical reports from patients with BI-RADS scores of 0, 1, or 2, designed to facilitate non-commercial research in breast cancer imaging.

The Gist

Imagine you are trying to find a specific needle in a haystack, but the hay is so thick and tangled that the needles hide behind each other. This is exactly the challenge radiologists face when looking at standard 2D mammograms of women with dense breast tissue. The layers of tissue overlap, making it hard to see if a "needle" (a potential cancer) is actually there or just a trick of the light.

Enter DBT-2026, a new "super-tool" for researchers, described in this paper. Here is the breakdown in simple terms:

1. The Problem: The "Flat" Picture vs. The "3D" Slice

Traditional mammograms are like taking a flat photograph of a sandwich. If you have too many ingredients (dense tissue), you can't tell if there's a hidden pickle (a tumor) or just a slice of tomato overlapping it.

Digital Breast Tomosynthesis (DBT) is like taking that sandwich and slicing it into thin layers, then looking at each slice individually. It's a 3D view that separates the overlapping tissue, making it much easier to spot the "pickles." While doctors are getting better at using this 3D tech, they need a massive library of examples to teach computers (Artificial Intelligence) how to spot these issues automatically.

2. The Solution: A Massive, Annotated "Training Library"

The authors created DBT-2026, which is essentially a giant, open-source library of 558 real-world 3D breast scans.

• The Collection: They gathered scans from 558 different women.
• The "Answer Key": What makes this special is that for every scan, they have the "answer key." They know for a fact if the patient had cancer, a benign lump, or nothing at all, because many of these women had biopsies (where a tiny piece of tissue was tested in a lab).
• The Privacy Shield: Just like you wouldn't want your medical records posted on a billboard, the researchers scrubbed every single piece of personal information (names, dates, locations) from the images and reports. They used advanced computer tools (like a digital eraser) to ensure no one could figure out who the patients were.

3. The "Team of Detectives"

You can't just dump a pile of photos on a computer and expect it to learn. Humans had to label them first.

• A team of highly trained breast imaging experts (like master detectives) looked at every single scan.
• They drew circles around suspicious spots, measured them, and wrote down exactly what they saw.
• They used a "Doer-Checker" system: One expert labeled the image, and a second expert double-checked their work to make sure it was perfect.
• Finally, a senior US-based radiologist gave the final "stamp of approval" to ensure the labels were accurate.

4. Who Can Use It? (The Rules of the Game)

The authors are making this library free for researchers, but with strict rules:

• Non-Commercial Only: You can use it to learn, teach, and invent new ways to detect cancer. You cannot sell it or use it to make a profit.
• No Clinical Use: You cannot use this specific dataset to diagnose a real patient in a hospital. It is for research and training AI models only.
• No Re-identification: You are strictly forbidden from trying to figure out who the patients were.

5. Why Does This Matter?

Think of AI as a student. To become a doctor, a student needs to study thousands of textbooks and practice cases. Before this dataset, AI students trying to learn 3D breast imaging had very few textbooks to study, and many of them were missing the "answer key" (biopsy results).

DBT-2026 is like handing that student a complete, high-quality textbook with the answers in the back. By giving researchers this data, the authors hope to speed up the development of AI tools that can help doctors find breast cancer earlier and more accurately, especially in women with dense breasts where it's hardest to see.

In a nutshell: They built a secure, 3D "training gym" for AI, filled with real cases and expert labels, so that future computer programs can become better at spotting breast cancer and saving lives.

Technical Summary

1. Problem Statement

• Clinical Context: Digital Breast Tomosynthesis (DBT), or 3D mammography, is increasingly adopted for breast cancer screening, particularly for women with dense breast tissue where 2D mammography suffers from tissue overlap. DBT improves lesion visibility, characterization, and localization.
• The Gap: While Artificial Intelligence (AI) has been extensively studied for 2D mammography, there is a significant scarcity of well-annotated, publicly available DBT datasets that include high-quality ground truth outcomes (specifically biopsy results). Existing public DBT datasets often lack sufficient numbers of malignant cases or detailed annotations, hindering the development and validation of robust AI models for 3D breast imaging.

2. Methodology

The authors curated the DBT-2026 dataset through a rigorous multi-step process involving data sourcing, de-identification, and expert annotation.

• Data Sourcing & Selection:

  • Source: DBT exams were retrieved from a hospital Picture Archiving and Communication System (PACS) in the US Midwest via Segmed's automated pipeline.
  • Inclusion Criteria: Female patients $\ge$ 18 years old eligible for screening; bilateral exams containing both Mediolateral Oblique (MLO) and Craniocaudal (CC) views; no breast implants, prior surgery, or history of breast cancer.
  • BI-RADS Filtering: The cohort was restricted to patients with Breast Imaging Reporting and Data System (BI-RADS) scores of 0, 1, or 2.
    • Score 0: Incomplete assessment (abnormality found, needs further imaging).
    • Score 1: Negative.
    • Score 2: Benign findings.
  • Enrichment Strategy: The dataset was specifically enriched by prioritizing patients with biopsy-proven malignancies to ensure a balanced distribution for AI training.

• De-identification & Privacy:

  • Process: Protected Health Information (PHI) was removed using a combination of Optical Character Recognition (OCR) for burnt-in text, Natural Language Processing (NLP), and Large Language Models (LLMs) to scrub the 18 HIPAA-defined identifiers from both images and free-text radiology reports.
  • Verification: A clinical specialist manually verified the de-identified images and text to ensure no PHI remained.

• Annotation Workflow:

  • Annotators: A team of fellowship-trained senior Breast Imaging Radiologists (based in India, recruited via iMerit) performed the annotations.
  • Protocol: A standardized protocol developed by US-based clinical experts was used for lesion localization, characterization, segmentation, and measurement.
  • Quality Control: A "Doer–Checker" workflow was implemented where every case was independently reviewed. Discrepancies were resolved via consensus. A final Quality Assurance (QA) review was conducted by a US-based MQSA-certified radiologist.
  • Tools: Annotations were performed using the HIPAA-compliant Ango Radiology Annotation Suite.

3. Key Contributions

• Dataset Release: Introduction of DBT-2026, a publicly available dataset containing 558 unique DBT exams from 558 patients.
• High-Quality Ground Truth: Unlike many public datasets, this collection includes biopsy-proven outcomes (ground truth) for a significant portion of the cohort, alongside expert lesion segmentation and characterization.
• Comprehensive Metadata: The dataset includes structured data (BI-RADS scores, biopsy results, demographics) and unstructured data (de-identified free-text radiology reports).
• Accessibility: The dataset is made freely available to researchers for non-commercial projects via a registration and credentialing process, with strict prohibitions on patient re-identification and clinical decision-making use.

4. Results & Dataset Statistics

The resulting dataset comprises 558 patients with a mean age of 57.5 ± 11.1 years. The demographic breakdown is predominantly White (537 patients), with a mean age of 57.5 years.

Cohort Distribution (Table 2):
The dataset is stratified into four distinct groups to facilitate diverse AI training scenarios:

1. Group A (Malignant): 271 patients (48.6% of cohort) with biopsy-proven malignancies.
2. Group B (Benign): 140 patients with benign findings.
3. Group C (Non-cancer Callback): 115 patients who were callbacks (BI-RADS 0) but did not have cancer.
4. Group D (Biopsy-proven Benign): 32 patients with benign biopsy results.

BI-RADS Distribution:

• BI-RADS 0: 325 exams
• BI-RADS 1: 91 exams
• BI-RADS 2: 142 exams

Total Biopsies: 538 biopsies were performed across the cohort, providing a robust ground truth for the majority of cases.

5. Significance

• AI Advancement: DBT-2026 addresses a critical bottleneck in medical AI by providing a high-quality, annotated 3D mammography dataset. This will enable the training and validation of AI models for breast cancer detection, characterization, and localization in DBT images.
• Clinical Relevance: By including a high proportion of biopsy-proven malignancies (nearly 50%) and specific BI-RADS 0 cases (where diagnostic uncertainty is highest), the dataset is uniquely suited for developing models that assist radiologists in difficult diagnostic scenarios.
• Research Community: The dataset lowers the barrier to entry for researchers working on breast imaging AI, fostering collaboration and accelerating the development of tools that could improve screening efficiency and reduce false-positive recalls.

Limitations:

• The data is sourced from a single hospital, which may limit generalizability across different populations or scanner manufacturers.
• Usage is restricted to non-commercial research only.