Breast-Lesion-USG Dataset Overview
- Breast-Lesion-USG is a curated 2D benchmark featuring expert lesion masks used for precise segmentation and uncertainty analysis.
- The dataset comprises 256 ultrasound cases processed as 512×512 patches, with in-domain metrics such as Dice 0.7726 and IoU 0.6801.
- It serves as both a state-of-the-art in-distribution benchmark and a stress test for evaluating cross-dataset generalization.
Searching arXiv for papers discussing the Breast-Lesion-USG dataset and closely related breast ultrasound datasets. The term Breast-Lesion-USG Dataset appears in recent breast-ultrasound literature in two closely related senses. In the most specific usage, it denotes a curated 2D breast ultrasonography benchmark with expert lesion masks that is cited as Pawłowska et al., “Curated benchmark dataset for ultrasound based breast lesion analysis,” and is used as a clean reference dataset for segmentation, uncertainty estimation, and cross-domain evaluation. In adjacent usage, closely related labels such as “BrEaST-Lesions USG Images and Masks” refer to segmentation-focused breast-ultrasound resources with expert-delineated masks, and BrEaST is also used as a source component in larger composite datasets for joint segmentation and classification (Musah et al., 25 Aug 2025, Abbadi et al., 5 Sep 2025, Johnny et al., 9 Jan 2026).
1. Nomenclature and dataset identity
In arXiv studies, Breast-Lesion-USG is most explicitly described as a curated benchmark for breast ultrasound lesion analysis. One segmentation-and-uncertainty study cites it directly as the dataset of Pawłowska et al. and uses it as the primary in-distribution benchmark for breast-ultrasound segmentation (Musah et al., 25 Aug 2025). A separate multi-dataset study refers to a closely related resource as “BrEaST-Lesions USG Images and Masks” and characterizes it as a set focused on lesion segmentation with high-quality expert-delineated masks, supporting classification and localization tasks, particularly in small-lesion scenarios (Abbadi et al., 5 Sep 2025).
A further extension appears in the PRECISE 2025 multi-task study, where BrEaST is one of the three public BUS sources merged into a composite breast-ultrasound dataset alongside BUSI and BUS-BRA. In that setting, BrEaST contributes curated images and masks rather than standing alone as the entire benchmark (Johnny et al., 9 Jan 2026). This suggests that “Breast-Lesion-USG” functions both as a specific benchmark label and as a shorthand for a family of curated breast-lesion ultrasound resources centered on lesion masks and lesion-level analysis.
2. Core data characteristics and annotation regime
The clearest dataset-level description is the one used for trustworthy segmentation and uncertainty estimation. There, 256 ultrasound cases are analyzed, comprising 83,099,921 total pixels and 5,814,303 foreground lesion pixels. Images are processed as standardized patches. The dataset provides expert manual segmentation masks for breast lesions, and these masks serve as the ground truth for supervised training and for evaluation with Dice and IoU. In that study, no deduplication or cleaning is reported for Breast-Lesion-USG; instead, it is treated as a clean benchmark in contrast to BUSI (Musah et al., 25 Aug 2025).
A related description in the transfer-learning literature emphasizes the same annotation orientation. BrEaST-Lesions USG Images and Masks is described as a segmentation-focused set with high-quality expert-delineated masks, and its role is explicitly tied to localization and small-lesion analysis rather than to weak labels or coarse boxes (Abbadi et al., 5 Sep 2025). In the PRECISE 2025 composite dataset, the BrEaST component contributes 99 benign and 63 malignant images, with 107 masks for 99 images because some images have multiple lesions (Johnny et al., 9 Jan 2026).
These descriptions establish the dataset family as fundamentally mask-centric. Unlike datasets organized only around benign/malignant labels, the Breast-Lesion-USG line is repeatedly used where pixel-level delineation is required.
3. Role as a segmentation benchmark
Breast-Lesion-USG is used as a benchmark for modified Residual Encoder U-Net segmentation within the nnUNet framework. In that configuration, the network has 8 encoder stages and 7 decoder stages, uses 2D patches, batch size 13, SGD, Batch Dice loss, and 250 epochs with 5-fold cross-validation (Musah et al., 25 Aug 2025).
The benchmark is important because it separates in-distribution performance from out-of-distribution transfer. When the model is trained and validated directly on Breast-Lesion-USG, it reaches a Dice coefficient of 0.7726 and an IoU of 0.6801. When trained on BUSI-A3 and tested out of domain on Breast-Lesion-USG, performance falls to Dice 0.4855 and IoU 0.4309 (Musah et al., 25 Aug 2025).
| Setting | Dice | IoU |
|---|---|---|
| In-distribution benchmark on Breast-Lesion-USG | 0.7726 | 0.6801 |
| BUSI-A3 trained, Breast-Lesion-USG tested OOD | 0.4855 | 0.4309 |
This benchmark use is methodologically significant because it makes Breast-Lesion-USG serve two roles simultaneously: a state-of-the-art in-domain target and a stress test for cross-dataset generalization. The reported gap quantifies a substantial domain-shift effect rather than a marginal variation in test conditions (Musah et al., 25 Aug 2025).
4. Uncertainty quantification, calibration, and trustworthiness
Breast-Lesion-USG is also central to recent work on epistemic uncertainty estimation in breast-ultrasound segmentation. All uncertainty analyses in that study are performed on the 256 Breast-Lesion-USG cases, covering over 83 million pixels. The evaluated approaches are MC dropout, deep ensembles, and a combined deep ensemble + MC dropout strategy. MC dropout uses stochastic forward passes, while the combined method uses 5 ensemble members × 3 MC samples per member = 15 samples per image (Musah et al., 25 Aug 2025).
The same work reports that deep ensembles and the combined method achieve slightly better ECE and pixel-wise accuracy than MC dropout alone, while uncertainty maps remain clinically interpretable: high-uncertainty regions align with lesion boundaries, noisy, artifact-heavy regions, and atypical or complex lesions. Under OOD testing on Breast-Lesion-USG, segmentation accuracy decreases and uncertainty values increase, which is precisely the desired failure-signaling behavior for trustworthy deployment (Musah et al., 25 Aug 2025).
The technical implication is not merely that uncertainty can be computed, but that Breast-Lesion-USG supplies sufficiently consistent masks and case structure to support pixel-level calibration analysis. In this usage, the dataset is not only a segmentation corpus but also a benchmark for whether uncertainty rises in regions where the model is unreliable.
5. Relation to composite datasets and to video-based breast-ultrasound resources
Breast-Lesion-USG should not be conflated with all breast-ultrasound datasets. In the PRECISE 2025 multi-task study, BrEaST is merged with BUSI and BUS-BRA into a 2,500-image composite dataset with 2,508 masks. There, BrEaST is one curated source among several and supports a prompt-free SAM-based framework for joint segmentation and classification rather than serving as the sole benchmark (Johnny et al., 9 Jan 2026).
It also should not be conflated with video datasets. Video-oriented breast-ultrasound resources address different tasks and have different annotation structures. Examples include a detection dataset of 188 ultrasound videos with 25,272 frames and frame-level rectangles plus video-level benign/malignant labels (Lin et al., 2022), a segmentation dataset of 572 videos with 34,300 annotated frames plus lesion masks, bounding boxes, benign/malignant labels, and ALN metastasis labels (Lin et al., 2023), and UVBLS200, which contains 200 ultrasound video sequences and 10,666 frames with per-frame lesion masks (Tu et al., 2024). These are temporally organized benchmarks for detection or video segmentation; Breast-Lesion-USG, by contrast, is used in the cited studies as a 2D curated lesion-mask benchmark.
A separate small but paired-modality resource is the B-US/SE-US dataset of 85 patients, 261 B-US images, and 261 SE-US images, retrospectively gathered for benign/malignant classification with biopsy confirmation (Misra et al., 2021). That dataset is important for multimodal CAD, but it is distinct from the curated Breast-Lesion-USG benchmark and does not play the same role in the segmentation literature.
6. Limitations, availability, and research significance
The literature treats Breast-Lesion-USG as curated and consistent, but it also exposes concrete limitations. A major one is foreground–background imbalance: in the uncertainty study, only 5,814,303 of 83,099,921 analyzed pixels are foreground lesion pixels, so pixel-wise accuracy can appear high partly because background dominates. Another is domain specificity: strong in-domain performance on Breast-Lesion-USG does not guarantee transfer from other BUS datasets, as the BUSI-A3-to-Breast-Lesion-USG drop in Dice from 0.7726 to 0.4855 demonstrates (Musah et al., 25 Aug 2025).
Availability is described unevenly across derivative studies. In the uncertainty-estimation work, the dataset is not redistributed by the paper itself and is instead referenced through the original publication (Musah et al., 25 Aug 2025). In the multi-dataset transfer-learning study, BrEaST-Lesions USG Images and Masks is treated as one of the publicly available breast-ultrasound datasets sourced from public academic repositories, but no direct URL is given there (Abbadi et al., 5 Sep 2025). This suggests that access is mediated by the original source rather than by downstream method papers.
The dataset’s significance lies in the combination of expert lesion masks, curated case structure, and repeated use as a reference domain. It anchors in-distribution segmentation benchmarks, makes OOD degradation measurable, and supports calibration studies in which uncertainty maps can be compared directly against expert contours. Within current arXiv literature, that combination gives Breast-Lesion-USG a role that is more specific than a generic breast-ultrasound collection and narrower than video or multimodal CAD datasets, but also more methodologically valuable for segmentation-centric research (Musah et al., 25 Aug 2025, Johnny et al., 9 Jan 2026).