BrEaST-Lesions USG: Ultrasound Breast Lesion Analysis

• BrEaST-Lesions USG is an ultrasound-based approach that integrates BI-RADS assessment, quantitative morphometrics, and segmentation to analyze breast lesions.
• It converts subjective descriptors like shape, margin, and echogenicity into objective, reproducible metrics that enhance specificity and reduce unnecessary biopsies.
• Recent advances leverage deep learning and multimodal fusion to improve lesion classification, segmentation accuracy, and overall risk estimation.

BrEaST-Lesions USG denotes the ultrasound-based study of breast lesions, spanning lesion depiction, BI-RADS assessment, quantitative morphometrics, segmentation, computer-aided diagnosis, and, in recent dataset usage, annotated image collections such as Breast Lesions USG or BLU. In this domain, ultrasound is valued because it is non-invasive, relatively inexpensive, widely available, and particularly useful in dense breasts, yet its interpretation remains strongly operator-dependent. Current research therefore focuses on converting subjective descriptors—shape, margins, orientation, echogenic pattern, and lesion extent—into reproducible computational representations that can improve specificity, preserve high sensitivity, and reduce unnecessary biopsies (Byra et al., 2017).

1. Clinical framing and diagnostic role

Breast ultrasound is used to detect, localize, and characterize masses, to complement mammography and MRI, and to guide biopsy and follow-up. A central clinical problem is that biopsy thresholds are intentionally conservative: lesions assigned BI-RADS 4 or 5 are typically referred for tissue sampling, which preserves sensitivity but yields many benign biopsies. In a study of 107 biopsy-proven solid lesions, all malignant lesions were assigned BI-RADS $\geq 4a$, so the threshold “BI-RADS $> 3$ considered malignant” produced 100% sensitivity but only 54.7% specificity and 68.2% accuracy (Byra et al., 2017).

This operator dependence is not confined to single-modality reading. A multimodal interface study involving 31 clinicians and 566 images showed that joint viewing of mammography, ultrasound, and MRI reduced BI-RADS error relative to single-modality reading, especially in the most severe case, and also reduced time spent per image in the multimodality condition (Calisto et al., 2020). A plausible implication is that ultrasound lesion analysis is most informative when treated as part of a structured, cross-modal diagnostic workflow rather than as an isolated image interpretation problem.

Recent multicenter work has pushed this framing further by treating breast ultrasound interpretation as two related but distinct tasks: biopsy recommendation and malignancy prediction. In 1,747 women across three centers, separate nomograms were built for these endpoints using BI-RADS descriptors, morphometric features, or both, showing that ultrasound lesion analysis increasingly functions as a calibrated risk-estimation problem rather than a purely categorical reading task (Ardakani et al., 31 Aug 2025).

2. Data resources, annotation, and the meaning of “BrEaST-Lesions USG”

A defining feature of the literature is the coexistence of pathology-grounded classification datasets and expert-mask segmentation datasets. The morphometric BI-RADS fusion study used 214 images from 107 solid lesions, acquired in two orthogonal planes per lesion and averaged across planes; pathology provided the ground truth, with 75 benign and 32 malignant lesions (Byra et al., 2017). Segmentation studies typically rely on manual lesion contours drawn by radiologists and then use those contours as pixel-level supervisory targets.

In current dataset terminology, Breast Lesions USG or BLU refers to a mask-annotated breast ultrasound dataset used for segmentation and prompt-conditioned modeling. In one recent study, BLU contained 256 breast ultrasound images from 256 patients; after excluding 4 images with more than one annotated lesion, 252 images remained, including 154 benign and 98 malignant tumors. The dataset provides pixel-wise masks together with structured metadata such as BI-RADS labels, lesion shape, margin, and histopathology (Mallina et al., 8 Sep 2025). A separate multi-dataset classification study describes BrEaST-Lesions USG Images and Masks as a resource emphasizing expert-delineated masks, lesion localization, and small-lesion scenarios, mainly for segmentation, augmentation, and external validation rather than primary headline classification results (Abbadi et al., 5 Sep 2025).

Ground-truth generation itself is a technical issue. In one collected ultrasound dataset for global-guidance segmentation, 632 images from 200 patients were manually delineated by three experienced radiologists, with final masks refined by a senior radiologist; in another multicenter nomogram study, lesion segmentations for morphometric extraction were performed manually by two radiologists and filtered by intraclass correlation coefficient before modeling (Xue et al., 2021, Ardakani et al., 31 Aug 2025). This suggests that annotation quality, not only model design, is a major determinant of reported performance.

3. Morphometric characterization and BI-RADS fusion

A major research line in BrEaST-Lesions USG is the conversion of lesion boundary appearance into explicit quantitative descriptors. In the 2017 study on the added value of morphological features, 30 lesion-shape features were extracted from ultrasound contours, including depth-width ratio, NRL entropy, circularity, elliptic-normalized circumference, normalized residual value, overlap ratio, extent, and morphological closing ratio. Representative definitions include circularity, often expressed as $\frac{4\pi A}{P^2}$, and extent, defined as lesion area divided by the area of the bounding rectangle. Convex-hull-derived measures such as overlap ratio and normalized residual value quantify non-convexity, spiculation, and contour deficit relative to a smooth envelope; NRL entropy summarizes irregularity in radial boundary variation through an entropy term of the form $H=-\sum_j p_j \log p_j$ (Byra et al., 2017).

The central empirical result of that study was that BI-RADS alone was not optimal. A logistic-regression classifier combining BI-RADS with six morphological features—extent, overlap ratio, NRL entropy, circularity, elliptic-normalized circumference, and normalized residual value—achieved an AUC of 0.986. At the operating point constrained to 100% sensitivity, specificity increased to 74.7% and accuracy to 82.2%, compared with 54.7% specificity and 68.2% accuracy for BI-RADS alone. In the same dataset, the combined model would have spared 56 of 75 benign biopsies while still detecting all malignant lesions (Byra et al., 2017).

This idea has since matured into interpretable nomograms. In the multicenter study of 1,747 women, 10 BI-RADS features and 26 morphometric characteristics were extracted, and separate BI-RADS, morphometric, and fused logistic nomograms were constructed. The general model form was the standard logistic expression

$$\log\frac{p}{1-p}=\beta_0+\sum_{j=1}^{k}\beta_j x_j.$$

For biopsy recommendation, the fused nomogram reached pooled accuracy 83.0% and AUC 0.901; for malignancy prediction, pooled accuracy was 83.8% and AUC 0.853. The fused model outperformed the morphometric nomogram, three radiologists, and two ChatGPT variants, while the BI-RADS nomogram alone also remained strong (Ardakani et al., 31 Aug 2025). This trajectory indicates that hand-crafted morphology has not disappeared in the deep-learning era; instead, it has become an interpretable quantitative complement to BI-RADS.

4. Segmentation paradigms in breast ultrasound

Automatic segmentation is the most intensively studied subproblem in breast ultrasound because shape, margin, size, and many radiomic descriptors depend directly on the lesion mask. Across the literature, the same failure modes recur: speckle noise, similar intensity distributions between lesion and background, blurred or indistinct margins, and especially irregular malignant contours (Chen et al., 2022, Chen et al., 2022, Yang et al., 2023).

Early deep adversarial approaches already showed strong overlap performance. SPCGAN, a semi-pixel-wise cycle GAN for 2D breast ultrasound, reported Dice 0.92 overall, compared with 0.90 for FCN and 0.79 for a level-set baseline; for malignant lesions, Dice improved from 0.90 with FCN to 0.93 with SPCGAN (Xing et al., 2019). Architectures then moved toward explicit context and edge modeling. GG-Net introduced global guidance blocks and boundary detection modules, reaching Dice 87.1% on a collected breast ultrasound dataset and 82.1% on BUSI, with consistent gains over U-Net, U-Net++, FPN, DeepLabv3+, AG-Unet, and DAF (Xue et al., 2021).

Attention-based encoder-decoder models further specialized this strategy. AAU-net replaced standard convolutions with a hybrid adaptive attention module combining channel and spatial self-attention across multiple receptive fields. On BUSI it reported Dice $77.51 \pm 0.68\%$, and in external validation on STU it reported Dice $80.23 \pm 2.60\%$, together with the lowest boundary distances among compared methods (Chen et al., 2022). PBNet later combined a multilevel global perception module with a boundary-guided module and a boundary-enhanced loss; the paper reports improvements over the state of the art by 0.70% Dice, 1.1% Jaccard, 0.1% specificity, and 2.5% HD95, with PBNet reaching Dice $81.10 \pm 2.21\%$ on BUSI and $83.11 \pm 1.96\%$ on an in-house dataset (Yang et al., 2023). For malignant-only segmentation, BAGNet explicitly targeted the more difficult malignant subset of BUS images and achieved Dice $69.93 \pm 3.63\%$ and Jaccard $> 3$0 on 210 malignant cases from BUSIS (Chen et al., 2022).

Weak and prompt-based supervision have recently expanded the design space. A weakly supervised pipeline combining morphology enhancement, CAM localization, and SAM reached Dice 74.39% on BUSI without any pixel-level labels during training and showed boundary accuracy comparable to supervised baselines by HD95 (Yue et al., 2023). On BLU, XBusNet introduced a dual-prompt, dual-branch multimodal architecture that couples a CLIP-based global branch with a U-Net-like local branch modulated by BI-RADS-informed text prompts. It achieved mean Dice 0.8765 and IoU 0.8149 under five-fold cross-validation, with the largest gains in small lesions (Mallina et al., 8 Sep 2025). This suggests that segmentation research in BrEaST-Lesions USG has shifted from purely image-driven contour extraction toward clinically conditioned, metadata-aware mask generation.

5. Classification, multimodality, and lesion-level representation learning

Breast lesion classification in ultrasound has also evolved beyond single-image benign/malignant prediction. A modified VGG16 with an attention block and an ensembled loss combining binary cross-entropy and log-cosh achieved 93% accuracy, 96% sensitivity, 90% specificity, and MCC 0.87 on a combined clinical and public ultrasound dataset, outperforming several modified VGG16 variants (Kalaf et al., 2021). This line of work emphasizes lesion-focused attention without explicit manual segmentation.

Another major direction is multimodal ultrasound fusion. A transfer-learning ensemble combining B-mode ultrasound and strain elastography used paired 4-channel inputs and a feature-level ensemble of AlexNet and ResNet-18. In the best patient-wise, cropped-lesion setting, the model achieved 90.00% accuracy, 88.89% sensitivity, and 91.10% specificity, exceeding manual strain-ratio baselines on the same dataset (Misra et al., 2021). This indicates that stiffness information and morphology are complementary, not redundant.

Beyond modality fusion, several studies treat multiple views of the same lesion as a supervision resource. A self-supervised lesion-recognition framework used lesion-level contrastive learning so that different views of the same lesion were pulled together in representation space. On 5,911 images from 2,131 lesions, the lesion-recognition auxiliary task improved AUC from 94.7 to 96.2 and accuracy from 86.3 to 89.1 relative to classification alone, while also reducing inner-lesion misclassification rate (Guo et al., 2022). A related multimodality study combined representative static ultrasound images with dynamic video, using the complementarity of radiologist-selected stills and real-time cine information; it reported 90.0% AUC and 81.7% accuracy on 897 image-video sets (Huang et al., 2023).

Current work has also moved beyond binary classification toward subtype recognition. A long-tailed classification framework for eight histological subtypes of breast ultrasound lesions used class-conditional latent diffusion synthesis, a sketch-grounded perception branch, and reinforcement-learning-based adaptive sampling. On the Breast-LT-8 dataset of 5,622 images from 2,811 patients, the full model increased F1 to 35.23 and few-shot accuracy to 31.25, whereas the baseline few-shot accuracy was 0.00 (Chen et al., 30 Jul 2025). A plausible implication is that future BrEaST-Lesions USG systems will need to solve not only malignancy discrimination but also rare-subtype recognition under clinically realistic imbalance.

6. Interpretation, workflow integration, and unresolved issues

Interpretability and workflow integration are now explicit research targets. In the multicenter nomogram study, the fused BI-RADS-morphometric model outperformed both radiologists and LLMs for biopsy recommendation and malignancy prediction, while remaining fully interpretable through point-based nomograms (Ardakani et al., 31 Aug 2025). In a broader multi-dataset transfer-learning study, Grad-CAM was used to highlight diagnostically relevant regions in ultrasound-based breast cancer detection, and BrEaST-Lesions USG was included as a segmentation- and localization-oriented supporting dataset within that pipeline (Abbadi et al., 5 Sep 2025). The field is therefore increasingly differentiating between interpretability derived from human-readable features and interpretability derived from saliency or activation maps.

Multimodal context is extending beyond morphology. A learned super-resolution contrast-enhanced ultrasound approach reconstructed microvasculature in vivo from three breast lesions and recovered distinct vascular patterns corresponding to fibroadenoma, cyst, and invasive ductal carcinoma. Fibroadenoma showed homogeneous internal vascularization, cyst showed peripheral vessels with an avascular center, and invasive ductal carcinoma showed an irregular mass with peripheral hypervascularity and low central vascular concentration (Bar-Shira et al., 2021). This suggests that microvascular phenotype may become a complementary descriptor layer for BrEaST-Lesions USG, especially where B-mode morphology is equivocal.

At the same time, limitations remain structurally similar across studies. Commonly reported constraints include single-center or single-scanner cohorts, manual or semi-manual segmentation, absence of external validation, and reader-specific BI-RADS assignment. The 2017 morphology study used one experienced radiologist and one scanner with internal leave-one-out validation only; several segmentation works rely on BUSI or small in-house cohorts; and XBusNet explicitly notes the lack of cross-dataset external validation because BLU is currently unusual in combining masks with rich structured metadata (Byra et al., 2017, Chen et al., 2022, Mallina et al., 8 Sep 2025). Objectively, the literature shows steady gains in AUC, Dice, and specificity, but it also shows that annotation protocols, dataset composition, and lesion prevalence still shape conclusions as strongly as architecture choice.

BrEaST-Lesions USG is therefore best understood not as a single algorithmic problem but as a layered research program: standardized lesion description through BI-RADS, quantitative morphometrics for interpretable risk, segmentation for precise geometry, multimodal fusion for richer phenotype, and emerging dataset designs—such as BLU—that embed both masks and structured clinical descriptors into the modeling loop.