Source-linked AI summary
Detection of masses and architectural distortions in digital breast tomosynthesis: a publicly available dataset of 5,060 patients and a deep learning model
Mateusz Buda, Ashirbani Saha, Ruth Walsh, Sujata Ghate, Nianyi Li, Albert Święcicki, Joseph Y. Lo, Maciej A. Mazurowski
TL;DR
Public, well-annotated datasets for developing and evaluating breast-screening AI are limited, particularly for DBT. The authors released an annotated DBT dataset and a single-phase detection model; the model reached 65% sensitivity at 2 false positives per breast.
Problem
Well-annotated, large-scale public datasets are limited for developing and evaluating breast-screening AI, particularly for DBT.
Source: §1 · p. 2Method
The authors curated a public DBT dataset with annotations for masses and architectural distortions and developed a single-phase detection model.
Source: §1 · p. 3, §4 · p. 11Results
65% sensitivity at 2 FP per breast was achieved on the test set for all test cases.
Source: §3 · p. 11Takeaways & Limitations
The dataset is presented as a challenging, realistic benchmark for future DBT detection methods.
Source: §4 · p. 12The model used only 124 cancer and 175 benign lesion boxes, with no pretraining on other datasets or similar modalities.
Source: §4 · p. 12
Abstract
from arXiv · showhide
Breast cancer screening is one of the most common radiological tasks with over 39 million exams performed each year. While breast cancer screening has been one of the most studied medical imaging applications of artificial intelligence, the development and evaluation of the algorithms are hindered due to the lack of well-annotated large-scale publicly available datasets. This is particularly an issue for digital breast tomosynthesis (DBT) which is a relatively new breast cancer screening modality. We have curated and made publicly available a large-scale dataset of digital breast tomosynthesis images. It contains 22,032 reconstructed DBT volumes belonging to 5,610 studies from 5,060 patients. This included four groups: (1) 5,129 normal studies, (2) 280 studies where additional imaging was needed but no biopsy was performed, (3) 112 benign biopsied studies, and (4) 89 studies with cancer. Our dataset included masses and architectural distortions which were annotated by two experienced radiologists. Additionally, we developed a single-phase deep learning detection model and tested it using our dataset to serve as a baseline for future research. Our model reached a sensitivity of 65% at 2 false positives per breast. Our large, diverse, and highly-curated dataset will facilitate development and evaluation of AI algorithms for breast cancer screening through providing data for training as well as common set of cases for model validation. The performance of the model developed in our study shows that the task remains challenging and will serve as a baseline for future model development.
1 Introduction
The paper addresses limited public, well-annotated medical imaging data and challenges in comparing mammography AI results. It presents a public DBT dataset and a single-phase detection model.
1 Introduction
- Public, well-annotated medical imaging data are scarce, limiting deep-learning training and transparency in performance evaluation.Mammography studies also vary in datasets, tasks, and evaluation strategies, hindering reliable comparison.Source: §1 · p. 2
- The authors curated and annotated more than 22,000 DBT volumes from 5,060 patients and made the dataset publicly available.DBT provides multiple cross-sectional slices for each breast rather than mammography projection images.Source: §1 · pp. 2–3
- They also developed a publicly available single-phase deep-learning model for DBT abnormality detection as a baseline or fine-tuning resource.The authors identify the limited number of positive locations as a major modeling challenge.Source: §1 · p. 3
2.1 Dataset
The dataset was assembled from Duke Health System DBT studies and organized into normal, actionable, benign-biopsy, and cancer groups. It was split by patient and annotated for masses and architectural distortions.
2.1 Dataset
- The cohort began with 16,802 Duke Health System studies collected from 2014 to 2018, then was selected into four study groups.Some studies were excluded for foreign objects or lesions biopsied on a different tomosynthesis exam.Source: §2.1 · pp. 3–4
- The dataset included 5,129 normal studies from 4,609 patients without abnormal findings, 280 actionable studies from 278 patients resulting in further imaging for a mass or architectural distortion, 112 biopsied benign studies, and 89 biopsied cancer studies.Source: §2.1 · pp. 3–5
2.1.1 Split into training, validation, and test sets
- The 22,032 volumes from 5,610 studies and 5,060 patients were split into training, validation, and test sets with no patient overlap.The test set had 460 studies from 418 patients; validation had 312 from 280 patients; training had 4,838 from 4,362 patients.Source: §2.1.1 · p. 6
2.1.2 Image annotation
- Two experienced breast radiologists annotated biopsied masses and architectural distortions with rectangular boxes on central slices.They produced 190 cancer-lesion boxes and 245 benign-lesion boxes; 336 boxes marked masses and 99 marked architectural distortions.Source: §2.1.2 · pp. 6–7
2.2 The baseline algorithm
The baseline uses a single-stage, slice-based detector, with preprocessing and postprocessing for DBT volumes. The authors searched network and loss configurations, then defined test-set evaluation using FROC.
2.2.1 Preprocessing
- Preprocessing window-leveled and downscaled each slice, removed skin, and segmented the breast by retaining the largest connected component.A 2×2 local mean filter reduced slice dimensions by a factor of two.Source: §2.2.1 · p. 7
2.2.2 Detection algorithm
- The detector divides slices into 96×96-pixel cells, each predicting one box and a confidence score, using a DenseNet-based fully convolutional network.It was trained on cropped positive slices; inference combined predictions across volume halves and applied breast-region filtering and non-maximum suppression.Source: §2.2.2 · pp. 7–8
2.2.3 Experiments
- The grid search trained 768 models across network sizes and four objectness losses to address lesion-background class imbalance.The losses were binary cross-entropy, weighted binary cross-entropy, focal loss, and reduced focal loss; the selected model maximized validation sensitivity at 2 FP per DBT volume.Source: §2.2.3 · pp. 8–9
2.2.4 Final model evaluation on the test set
- Final evaluation used FROC, counting a detection as true positive when its center met the ground-truth proximity and slice-span criteria.The authors also assessed breast-based FROC, where any lesion detected on any view counted as a breast-level true positive.Source: §2.2.4 · p. 9
3 Results
On the test set, the selected model reached 65% sensitivity at 2 false positives per breast across all cases. Validation and test results also show performance differences across loss functions, evaluation levels, and case groups.
3 Results
- All tested loss functions performed similarly in 2D per-slice validation, with each loss's best configuration exceeding 78% sensitivity at 2 FP per slice.Figure 3 summarizes the loss-function comparison.Source: §3 · p. 9, Figure 3 · p. 10
- 60% sensitivity at 2 FP per DBT volume was the best validation result, achieved with focal loss; binary and weighted binary cross-entropy reached 59%, and reduced focal loss reached 58%.The focal-loss model was selected for test-set evaluation.Source: §3 · p. 10
- 42% sensitivity at 2 FP per DBT volume was achieved on the test set, with better performance on cancer than benign cases.Figure 4 shows the validation and test FROC curves for the selected model.Source: Figure 4 · p. 10, §3 · p. 11
- 65% sensitivity at 2 FP per breast was achieved across all test cases, while sensitivity for cancers was 67%.The breast-based test-set FROC is shown in Figure 5.Source: §3 · p. 11, Figure 5 · p. 11
4 Discussion
The study shares a large DBT dataset and a single-phase detection model as a baseline. Its inclusive screening cohort and evaluation design make the dataset a challenging, realistic benchmark, though differences from prior studies prevent direct comparison.
4 Discussion
- The publicly shared DBT dataset covers 5,060 patients, and the study trained a single-phase detection model as a baseline for future development.The dataset includes annotations for masses and architectural distortions, abnormalities that appear different in DBT images.Source: §4 · p. 11
- Including masses and architectural distortions, along with normal and actionable cases, makes the dataset more representative of screening practice and more challenging for automated detection.The authors describe normal and actionable cases as representative of a screening cohort.Source: §4 · pp. 11–12
- The model used 124 cancer and 175 benign lesion boxes without pretraining; it achieved moderate performance, whereas a prior from-scratch DBT-only model reported below 20% sensitivity at 2 FP per volume.The study's detector is a single-phase network that does not require multiple prediction-generation steps.Source: §4 · p. 12
- The evaluation is robust to large predicted boxes, but the 3D detection task generates more false positives than 2D detection tasks.Differences in abnormalities, case exclusions, and evaluation metrics make direct comparison with prior studies [21] virtually impossible.Source: §4 · p. 12