Breast cancer tumor detection and segmentation in ultrasound imaging is crucial for improving early diagnosis and guiding treatment decisions. This paper presents a two-stage deep learning framework that integrates lesion localization-guided semantic segmentation to enhance tumor detection in breast ultrasound images. In the first stage, YOLOv11-n-nano, a lightweight detection model with approximately 2 million parameters, is employed to localize suspicious regions of interest. In the second stage, advanced segmentation architectures-including U-Net++, U-Net3+, Attention U-Net, and TransUNet-are applied to refine the segmentation of the localized areas. The proposed framework is evaluated on the publicly available BUS_UCLM dataset. Our approach achieves state-of-the-art performance, with DeepLabV3 obtaining a Dice score of 94.31% and an Intersection over Union (IoU) of 92.10%. These results highlight the efectiveness of localization-guided segmentation in reducing false positives and improving segmentation accuracy in breast ultrasound imaging.
According to the World Health Organization, breast cancer is the most commonly diagnosed cancer
among women worldwide, with over 2.3 million new cases in 2020 and accounting for approximately
685,000 deaths annually [
Breast ultrasound imaging ofers several advantages: it is non-invasive, cost-efective, portable,
and free of ionizing radiation. However, ultrasound sufers from inherent limitations such as speckle
noise, low contrast, operator dependency, and variability in acquisition protocols [
In recent years, deep learning techniques, especially convolutional neural networks (CNNs), have
demonstrated remarkable success in medical image analysis tasks including classification, detection, and
segmentation [
To address these challenges, incorporating a lesion localization step prior to segmentation has been
proposed as an efective strategy. By guiding the segmentation model toward regions of interest, this
two-stage approach reduces false positives and improves delineation of lesion boundaries [
In this study, we present a two-stage deep learning pipeline for breast lesion analysis in ultrasound images. The first stage localizes suspicious regions using a YOLO-based model, while the second stage performs precise semantic segmentation with advanced architectures. We evaluate our framework on the BUS_UCLM dataset, a publicly available breast ultrasound lesion segmentation dataset containing 683 images across benign, malignant, and normal categories. We integrate four state-of-the-art CNNbased segmentation models: U-Net, U-Net++, TransUNet, Attention U-Net and DeepLab. We provide a comprehensive comparison of their performance.
The structure of this paper is as follows: Section 2 reviews related literature on breast tumor analysis in ultrasound imaging, with emphasis on deep learning-based methodologies. Section 3 introduces the proposed methodology, including the two-stage localization-segmentation framework, dataset preprocessing, and network architectures. Section 4 introduces the evaluation metrics and loss functions employed to assess the performance of the diferent models. Section 5 presents the experimental results and comparisons across diferent models. Finally, Section 6 concludes the paper with key findings, limitations, and future research directions.
Several studies have explored deep learning approaches for breast cancer detection and segmentation in ultrasound imaging.
Cho et al. [
Raza et al. [14] introduced DeepBreastCancerNet, a 24-layer CNN with inception modules, enhanced through transfer learning from nine pre-trained networks. Using the BUSI dataset (780 images) and an additional set of 250 images, the model achieved remarkable accuracy (99.35% and 99.63%), surpassing ResNet-50 (98.06%) and GoogLeNet (98.71%).
Mukasheva et al. [15] compared five UNet variants (UNet, Attention UNet, UNet++, DenseInception UNet, Residual UNet) on 780 ultrasound images. DenseInception UNet achieved the highest Dice (0.976) after augmentation, while Attention UNet underperformed, highlighting the influence of noise and preprocessing strategies.
Khaledyan et al. [16] studied UNet-based architectures with preprocessing (CLAHE, augmentation) on the BUSI dataset. Their proposed Sharp Attention UNet achieved Dice 0.93 and accuracy 97.9%, outperforming UNet, Sharp UNet, and Attention UNet.
MohammadiNasab et al. [17] presented a self-supervised multi-task approach (DATTR2U-Net) for Automated Breast Ultrasound (ABUS). Using inpainting and denoising as pretext tasks, their model improved segmentation robustness. On TDSC-ABUS, it outperformed Faster R-CNN and YOLO, though challenges remain for detecting very small lesions.
Madhu et al. [18] proposed UCapsNet, a two-stage model combining U-Net for segmentation and Capsule Networks for classification. On BUSI, it achieved outstanding results (segmentation Dice 99.07%, classification accuracy 99.22%), surpassing pre-trained CNNs like VGG-19 and ResNet-50.
Almajalid et al. [19] used an enhanced U-Net with preprocessing (denoising, contrast adjustment) and heavy augmentation on 221 images, achieving a Dice score of 82.5%, outperforming other automatic methods.
Vallez et al. [20] introduced the BUS-UCLM dataset (683 annotated ultrasound images). In baseline tests, Mask R-CNN performed best with Dice 77.09% and IoU 65.46%.
Zhang et al. [21] proposed a dual-branch DenseNet-UNet combining classification and segmentation on 1600 images, reaching AUC 0.991 and Dice 89.8%, and showing good generalization on external data.
Vakanski et al. [22] integrated visual attention maps into a modified UNet for 510 images, achieving Dice 90.5%, demonstrating the benefit of medical prior knowledge in guiding segmentation. Pan et al [23] introduces a difusion-based framework for synthesizing breast ultrasound images, integrating text prompts and a shape-aware mask generator to enhance realism and diversity. It efectively addresses data scarcity, with downstream evaluations showing strong performance boosts. On the BUS_UCLM dataset, segmentation achieves a peak DSC of 83.5% (Attention UNet at 25% synthetic ratio). Vallez et al. [24] evaluates deep learning approaches for breast ultrasound lesion detection and classification using the BUS_UCLM dataset. Among the models tested, Sk-UNet achieved the highest performance for multi-class segmentation with a IoU score of 79.2% and Dice of 87.3%.
Table 1 summarizes these related works.
Cho et al. [
Net
IoU = 0.77 ; Dice = 0.85 Precision = 0.99 Recall = 0.79 ; Precision = 0.56 ; SF = 0.65 IoU = 0.94 ; Dice = 0.95 ; Precision = 0.92 Dice = 0.82 Dice = 0.835 Poorya et al. [17]
Acc = 0.99 Almajalid et al. [19] 221 images
U-Net Vallez et al. [20]
Dice = 0.77 ; IoU = 0.65 Pan et al. [23]
This section presents two deep learning strategies developed for the automatic segmentation of breast lesions in ultrasound images. The goal is to enhance diagnostic accuracy and robustness by leveraging both direct and localization-guided segmentation approaches.
- Approach 1: Direct semantic segmentation using deep neural networks such as U-Net++,
U-Net3+, Attention U-Net, TransUNet, and DeepLabv3. - Approach 2: A localization-guided two-stage segmentation where YOLOv11-n is used to localize regions of interest, which are subsequently segmented using the same set of models.
Both approaches were implemented and evaluated on the BUS_UCLM dataset (Breast Ultrasound Lesion Segmentation Dataset) [20]. 3.1. Approach 1: Direct Semantic Segmentation This approach applies semantic segmentation directly to full breast ultrasound images, aiming to produce pixel-wise boundaries of tumoral regions. The objective is to generate precise binary masks that distinguish between lesion and non-lesion areas without relying on any prior localization step.
To evaluate this strategy, we implemented and compared five state-of-the-art segmentation models widely used in medical imaging: U-Net++ [25], U-Net3+ [26], Attention U-Net [27], TransUNet [28], and DeepLabv3 [29]. Each model was trained on the BUS_UCLM dataset and evaluated using standard performance metrics. The results obtained serve as a baseline for comparison with localization-guided segmentation strategies (see Figure 1). In the second approach, we introduce a two-step pipeline to refine segmentation by concentrating on the region of interest (ROI) of the lesion area identified in the original image. This strategy aims to reduce false positives and improve precision.
- Lesion localization: The YOLOv11-n-nano (YOLOv11-n) model, with around 2 million parameters, is a lightweight detector that delivers accurate results with low computational cost and fast inference. YOLOv11-n is applied to the full ultrasound image to localize bounding boxes around suspicious lesions. - Targeted Segmentation: Detected regions are cropped, resized, and passed to a segmentation model (same as those used in Approach 1) for semantic mask prediction. The resulting mask is then reintegrated into the original image coordinate space.
This guided strategy helps the model concentrate on lesion-relevant pixels, potentially improving performance in noisy or low-contrast ultrasound contexts (see Figure 1).
The localization-guided inference process is summarized in Algorithm 1. Algorithm 1: Localization-Guided Segmentation Pipeline using YOLOv11-n
Input: Preprocessed dataset (ultrasound images , ground truth masks )
YOLOv11-n detector segmentation model
Output: Segmentation metrics (Dice, IoU, etc.) 1 foreach ∈ do 2 = 11 − () // detect region(s) of interest; 3 if no localization then 4 Predict empty mask; 5 else 1. = Crop image using predicted bounding box; 2. = () //Apply segmentation model to obtain binary lesion mask; 3. Reproject segmented mask into original image dimensions; 4. Compute [Dice, IoU, Precision, Sensitivity, Recall](, ); 6 Return: [Dice, IoU, Precision, Sensitivity, Recall]; 3.3. BUS_UCLM Dataset The dataset we selected for ultrasound images in our approaches, BUS_UCLM (Breast ultrasound lesion segmentation dataset) [20], contains a total of 683 images acquired from 38 patients, divided into three categories: benign (174), malignant (90), and normal (419).
We split the dataset into 80% for the training set and 20% for the test set.
3.4. Preprocessing of the BUS_UCLM Dataset
Data preprocessing is an essential step to ensure the quality and consistency of the dataset, while also
improving model performance and reducing the risk of overfitting. For the BUS_UCLM dataset, which
consists of breast ultrasound images with corresponding lesion annotations, we applied the following
preprocessing steps:
• Resizing: All images were resized to a fixed resolution of 224 × 224 pixels, which is a common
requirement for convolutional neural network (CNN) architectures.
• Normalization: Pixel intensity values were normalized to the range [
To assess the efectiveness of our proposed approach for breast cancer diagnosis, we employed a set of standard evaluation metrics, including accuracy, F1-score, sensitivity, and precision, which are defined below. These metrics were used to quantify the performance of both the detection and segmentation components of the pipeline. In addition, appropriate loss functions were selected to guide model optimization during training. 4.1. Classification Metrics - Accuracy:
Accuracy =
+ + + +
Precision = + - Recall (also called Sensitivity or True Positive Rate):
Recall = +
Represents the counts of correct and incorrect classifications for binary classification problems. 4.2. Segmentation Metrics - Intersection over Union (IoU): - Dice Coeficient (F1 Score) : - Relation between Dice and IoU:
Dice = 12 +·IIooUU 4.3. Segmentation Loss Functions - Binary Cross Entropy (BCE):
IoU(, ) = | ∩ | = | ∪ |
+ + Dice(, ) = 2| ∩ | = || + ||
2 2 + + and
IoU =
Dice 2 − Dice 1 ∑︁ [ log() + (1 − ) log(1 − )] ℒBCE = − =1 where is the ground truth label, the predicted probability. - Dice Loss (derived from Dice coeficient): 2 ∑︀ ℒDice = 1 − ∑︀ + ∑︀ - Focal Loss (to address class imbalance):
ℒFocal = − ∑︁(1 − ) · log()
=1 where is the focusing parameter, and is the predicted probability. 4.4. YOLO Evaluation Metrics - Average Precision (AP): - Mean Average Precision (mAP):
AP =
Precision() ∫︁ 1
0 mAP =
1 ∑︁ AP =1 1 0.95 ∑︁
5.1. Approach 1: Direct Semantic Segmentation In the first strategy, segmentation models were directly applied to all ultrasound images of the BUS_UCLM dataset. We evaluated five architectures: UNet++, UNet3+, Attention U-Net, TransUNet, and DeepLabV3.
Table 3 summarizes the obtained results across diferent metrics (Loss, Dice, IoU, Recall, and Precision). As illustrated, DeepLabV3 significantly outperformed other models, achieving the highest Dice (84.49%) and IoU (73.57%), confirming its strong capability for tumor delineation.
The results demonstrate that DeepLabV3 clearly outperforms the other segmentation models in terms of Dice and IoU, which confirms its ability to capture fine tumor boundaries more efectively. UNet++ also shows competitive performance with a balanced precision and recall, making it a reliable alternative when computational eficiency is considered. In contrast, UNet3+ and Attention U-Net perform moderately, with Attention U-Net providing slightly higher recall but lower precision, indicating a tendency to over-segment. TransUNet achieved the weakest results, suggesting that transformer-based representations may require larger training datasets or additional fine-tuning to handle ultrasoundspecific noise and variability. Overall, the direct segmentation approach proves efective, but the variability in performance across models highlights the importance of architecture choice for breast ultrasound analysis.
Figure 3 presents visual examples of segmentation masks generated by the diferent models, highlighting the superior precision of DeepLabV3 compared to the other architectures. 5.2. Approach 2: Localization-Guided Semantic Segmentation The second approach was designed as a two-step process: (1) lesion detection using YOLOv11, followed by (2) focused segmentation on detected regions with the same five architectures.
YOLOv11 achieved competitive detection performance with an mAP@50 of 81.8% and mAP@50-95 of 65%, ensuring robust region proposals for the subsequent segmentation stage.
81.8% 65%
DeepLabV3 obtained the best overall results, achieving the highest Dice (94.31%) and IoU (92.10%), confirming its robustness for accurate lesion delineation.
TransUNet also performed competitively, recording the lowest loss (0.3061) and the highest recall (94.93%), which suggests a strong sensitivity in detecting lesion areas. In contrast, Attention U-Net achieved the best precision (94.67%), reflecting its ability to reduce false positives while maintaining high segmentation accuracy. Both UNet++ and UNet3+ yielded stable and reliable performances, with Dice scores of 94.27% and 94.08% respectively, further confirming their efectiveness in this task.
Figure 5 illustrates segmentation masks generated by the two-stage approach. Compared to the direct segmentation, the contours are more accurate, and false positives are reduced, confirming the benefit of integrating a detection step before segmentation. 5.3. Discussion Table 6 summarizes the performance of the two segmentation strategies. The direct segmentation approach (Approach 1) provides reasonable results, with DeepLabV3 achieving the highest Dice (84.49%) and IoU (73.57%), demonstrating its capability to capture tumor boundaries more accurately than the other architectures. UNet++ also shows competitive performance, while TransUNet and Attention U-Net are slightly weaker, either in precision or recall.
In contrast, the two-stage, YOLOv11-guided segmentation approach (Approach 2) substantially improves performance across all models. By first localizing lesion regions, the models can focus their learning on relevant areas, resulting in consistently higher Dice and IoU scores (all above 91%). DeepLabV3 again achieves the best overall performance (Dice 94.31%, IoU 92.10%), confirming its robustness for precise tumor delineation. Notably, TransUNet attains the highest recall (94.93%), indicating strong sensitivity to lesion regions, while Attention U-Net reaches the highest precision (94.67%), reflecting efective reduction of false positives.
Overall, the results highlight the benefits of integrating a detection step prior to segmentation. The two-stage framework not only enhances mask accuracy but also reduces over-segmentation and false positives, making it a reliable strategy for breast ultrasound analysis. 5.4. Comparison with Related Work The results clearly show that our approach substantially outperforms previous methods. While the models from Vallez et al. [20] reach only 68–77% in Dice and around 56–65% in IoU, all of our models exceed 94% in Dice and 91% in IoU, reflecting far more precise and reliable segmentation. Pan et al [ 23] 83.5% in Dice with Attention UNet at 25% synthetic ratio.
Among our models, DeepLabV3 slightly stands out with the best overall scores, but all others (UNet++, UNet3+, Attention UNet, and TransUNet) perform very closely, confirming the robustness of the proposed approach. In summary, our method brings a significant advancement in breast ultrasound image segmentation by efectively combining detection and segmentation.
Breast tumor segmentation in ultrasound imaging is essential for enhancing early diagnosis and improving patient management. In this study, we introduced a two-stage deep learning framework that integrates lesion localization with semantic segmentation to better delineate tumors in breast ultrasound images. The pipeline uses YOLOv11-n-nano, a lightweight detector of around 2 million parameters, for eficient and robust lesion localization, followed by leading segmentation architectures—U-Net3+, TransUNet, U-Net++,DeepLabV3, and Attention U-Net—to refine tumor boundaries. Evaluations on the BUS_UCLM dataset demonstrate that our approach significantly outperforms conventional methods.
The localization-guided segmentation strategy boosts precision by focusing the segmentation models on tumor-relevant areas, reducing false positives and emphasizing critical structures. Among the models tested, DeepLabv3 achieved the highest performance, with a Dice score of 94.31% and an IoU of 92.10%, while TransUNet’s strong results underscore the potential of hybrid convolutional-transformer designs. The proposed two-stage framework significantly enhances breast ultrasound lesion segmentation, achieving a Dice score of 94.31% and reducing diagnostic variability by minimizing manual delineations.
Looking ahead, future work will aim to enhance the clinical adaptability and generalization of this framework through several directions: • Extending the pipeline to process full ultrasound sequences or volumes, allowing the model to capture contextual continuity between consecutive frames. • Integrating multimodal ultrasound data—such as Doppler or elastography—to enrich lesion characterization and improve segmentation accuracy. • Validating the model across additional ultrasound datasets and acquisition protocols to ensure robustness and broad applicability. • Collaborating with clinicians for user-centered evaluations, assessing the tool’s utility in routine diagnostic workflows. • Optimizing the pipeline for real-time inference and deploying it on edge devices to support point-of-care ultrasound applications.
During the preparation of this work, the author(s) used ChatGPT in order to: Grammar and spelling check, Paraphrase and reword. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content. ultrasound images for breast cancer diagnosis, Journal of King Saud University-Computer and Information Sciences 34 (2022) 10273–10292. [14] A. Raza, N. Ullah, J. A. Khan, M. Assam, A. Guzzo, H. Aljuaid, DeepBreastCancerNet: A novel deep learning model for breast cancer detection using ultrasound images, Applied Sciences 13 (2023) 2082. [15] A. Mukasheva, D. Koishiyeva, Z. Suimenbayeva, S. Rakhmetulayeva, A. Bolshibayeva, G. Sadikova, Comparison evaluation of unet-based models with noise augmentation for breast cancer segmentation on ultrasound images., Eastern-European Journal of Enterprise Technologies 125 (2023). [16] D. Khaledyan, T. J. Marini, T. M. Baran, A. O’Connell, K. Parker, Enhancing breast ultrasound segmentation through fine-tuning and optimization techniques: Sharp attention unet, Plos one 18 (2023) e0289195. [17] P. MohammadiNasab, A. Khakbaz, H. Behnam, E. Kozegar, M. Soryani, A multi-task self-supervised approach for mass detection in automated breast ultrasound using double attention recurrent residual u-net, Computers in Biology and Medicine 188 (2025) 109829. [18] G. Madhu, A. M. Bonasi, S. Kautish, A. S. Almazyad, A. W. Mohamed, F. Werner, M. Hosseinzadeh, M. Shokouhifar, Ucapsnet: A two-stage deep learning model using u-net and capsule network for breast cancer segmentation and classification in ultrasound imaging, Cancers 16 (2024). [19] R. Almajalid, J. Shan, Y. Du, M. Zhang, Development of a deep-learning-based method for breast ultrasound image segmentation, in: 2018 17th IEEE International Conference on Machine Learning and Applications (ICMLA), IEEE, 2018, pp. 1103–1108. [20] N. Vallez, G. Bueno, O. Deniz, M. A. Rienda, C. Pastor, Bus-uclm: Breast ultrasound lesion segmentation dataset, Scientific Data 12 (2025) 242. [21] S. Zhang, M. Liao, J. Wang, Y. Zhu, Y. Zhang, J. Zhang, R. Zheng, L. Lv, D. Zhu, H. Chen, W. Wang, Fully automatic tumor segmentation of breast ultrasound images with deep learning, Journal of Applied Clinical Medical Physics 24 (2023) e13863. URL: https://doi.org/10.1002/acm2.13863. doi:10.1002/acm2.13863. [22] A. Vakanski, S. G. Lee, N. Baker, Attention-based segmentation of breast tumors in ultrasound images using saliency maps, Sensors 20 (2020) 2612. doi:10.3390/s20092612. [23] H. Pan, H. Lin, Z. Feng, C. Lin, J. Mo, C. Zhang, Z. Wu, Y. Wang, Q. Zheng, Breast ultrasound tumor generation via mask generator and text-guided network: A clinically controllable framework with downstream evaluation, arXiv preprint arXiv:2507.07721 (2025). [24] N. Vallez, I. Mateos-Aparicio-Ruiz, M. A. Rienda, O. Deniz, G. Bueno, Comparative analysis of deep learning methods for breast ultrasound lesion detection and classification, Physica Medica 134 (2025) 104993. [25] Z. Zhou, M. M. R. Siddiquee, N. Tajbakhsh, J. Liang, Unet++: A nested u-net architecture for medical image segmentation, in: N. Navab, J. Hornegger, W. Wells, A. Frangi (Eds.), Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support, Springer, 2018, pp. 3–11. [26] H. Huang, et al., Unet 3+: A full-scale connected u-net for medical image segmentation, in: ICASSP 2020 – IEEE International Conference on Acoustics, Speech and Signal Processing, IEEE, 2020, pp. 1055–1059. [27] O. Oktay, J. Schlemper, L. L. Folgoc, M. Lee, M. Heinrich, K. Misawa, K. Mori, S. McDonagh, N. Y.
Hammerla, B. Kainz, et al., Attention u-net: Learning where to look for the pancreas, arXiv preprint arXiv:1804.03999 (2018). [28] J. Chen, Y. Lu, Q. Yu, X. Luo, E. Adeli, Y. Wang, L. Lu, A. L. Yuille, Y. Zhou, Transunet: Transformers make strong encoders for medical image segmentation, arXiv preprint arXiv:2102.04306 (2021). [29] Y. Tang, D. Tan, H. Li, M. Zhu, X. Li, X. Wang, J. Wang, Z. Wang, C. Gao, J. Wang, et al., Rtc_tonguenet: An improved tongue image segmentation model based on deeplabv3, Digital Health 10 (2024) 20552076241242773.