PBE-UNet: A light weight Progressive Boundary-Enhanced U-Net with Scale-Aware Aggregation for Ultrasound Image Segmentation

Chen Wang; Yixin Zhu; Yongbin Zhu; Fengyuan Shi; Qi Li; Jun Wang; Zuozhu Liu; Keli Hu

cs.CVApr 15, 2026

PBE-UNet: A light weight Progressive Boundary-Enhanced U-Net with Scale-Aware Aggregation for Ultrasound Image Segmentation

Authors: Chen Wang, Yixin Zhu, Yongbin Zhu, Fengyuan Shi, Qi Li, Jun Wang, Zuozhu Liu, Keli Hu

Organizations: Department of Computer Science, Shaoxing University, Shaoxing, China · National Frontiers Science Center for Industrial Intelligence and Systems Optimization, Northeastern University, China · College of Computer Science, Chongqing University, China · School of Computer and Computing Science, Hangzhou City University, Hangzhou, China ·

Abstract

Accurate lesion segmentation in ultrasound images is essential for preventive screening and clinical diagnosis, yet remains challenging due to low contrast, blurry boundaries, and significant scale variations. Although existing deep learning-based methods have achieved remarkable performance, these methods still struggle with scale variations and indistinct tumor boundaries. To address these challenges, we propose a progressive boundary enhanced U-Net (PBE-UNet). Specially, we first introduce a scale-aware aggregation module (SAAM) that dynamically adjusts its receptive field to capture robust multi-scale contextual information. Then, we propose a boundary-guided feature enhancement (BGFE) module to enhance the feature representations. We find that there are large gaps between the narrow boundary and the wide segmentation error areas. Unlike existing methods that treat boundaries as static masks, the BGFE module progressively expands the narrow boundary prediction into broader spatial attention maps. Thus, broader spatial attention maps could effectively cover the wider segmentation error regions and enhance the model's focus on these challenging areas. We conduct expensive experiments on four benchmark ultrasound datasets, BUSI, Dataset B, TN3K, and BP. The experimental results how that our proposed PBE-UNet outperforms state-of-the-art ultrasound image segmentation methods. The code is at https://github.com/cruelMouth/PBE-UNet.

Figures & tables

Fig. 1: Visualization of prediction map, GT boundary, error map, and the boundary attention map generated by our method. (a) original image, (b) GT, (c) prediction map of U-Net, (d) GT boundary, (e) error map, (f) boundary attention map generated by our method. We can find that the narrow boundary fails to cover the wider error region, highlighting the limitation of simple feature fusion methods, such as “Add”, “Multiply”, and “Contact”. In contrast, our BGFE module addresses this issue by adaptively expanding the boundary into a broader attention region, which effectively aligns with and covers the error areas.

Fig. 2: The overall framework of the PBE-UNet network. It consists of the encoder, decoder, boundary detection (BD) module, boundary-guided feature enhancement (BGFE) module, and scale-aware aggregation module (SAAM).

Fig. 3: The architecture of boundary detection module.

Fig. 4: The architecture of boundary-guided feature enhancement module. First, we will concatenate the input features and the boundary together, and pass the fused features into a $3\times 3$ convolution network. Then, we operate multi-scale depth-wise convolutions for the boundary to expand the boundary information. Finally, we combine the expanded boundary and the enhanced features together through dot product operation.

Fig. 5: The architecture of scale-aware aggregation module (SAAM). We split the input features into four groups along the channel dimension. Then, we put the four groups into depth-wise convolutions with different dilation rates. Then, we concatenate the multi-scale features and utilize a ECA module to adaptively adjust the representations.

Fig. 6: Visual comparison of segmentation results of the proposed model with seven state-of-the-art methods on the BUSI dataset. (a) original image, (b) GT, (c) ours, (d) U-Net, (e) TransUNet, (f) CMUNet, (g) CMUNeXt, (h) GGNet, (i) BGNet, (j) BFNet.

Model	Dice $\%\color[rgb]{1,0,0}{\uparrow}$	IoU $\%\color[rgb]{1,0,0}{\uparrow}$	HD95 $\color[rgb]{1,0,0}{\downarrow}$	Recall $\%\color[rgb]{1,0,0}{\uparrow}$	Acc $\%\color[rgb]{1,0,0}{\uparrow}$
Traditional Segmentation Methods
U-Net (MICCAI2015) [ 23 ]	75.57	67.38	23.61	77.21	96.10
FCN-8s (CVPR2015) [ 16 ]	79.51	70.58	20.50	81.24	96.01
DeeplabV3+ (ECCV2018) [ 7 ]	81.31	73.19	15.73	84.30	96.25
AttUNet (MIDL2018) [ 20 ]	79.59	71.10	21.44	81.77	95.85
U-Net++ (TMI2020) [ 43 ]	79.11	70.32	24.04	83.06	96.87

TABLE I: Results on the BUSI dataset.

Traditional Segmentation Methods
Model	Dice $\%\color[rgb]{1,0,0}{\uparrow}$	IoU $\%\color[rgb]{1,0,0}{\uparrow}$	HD95 $\color[rgb]{1,0,0}{\downarrow}$	Recall $\%\color[rgb]{1,0,0}{\uparrow}$	Acc $\%\color[rgb]{1,0,0}{\uparrow}$
U-Net (MICCAI2015) [ 23 ]	80.29	72.90	14.69	78.99	98.44
AttUNet (MIDL2018) [ 20 ]	79.30	72.93	14.44	76.15	97.85
U-Net++ (TMI2020) [ 43 ]	80.73	71.04	14.80	78.26	97.97
TransUNet (ARXIV2021) [ 6 ]	83.77	75.28	12.56	84.11	97.92
Ultrasound Segmentation Methods

TABLE II: Results on the Dataset B dataset.

Fig. 7: Visual comparison of segmentation results of the proposed model with seven state-of-the-art methods on the Dataset B dataset. (a) original image, (b) GT, (c) ours, (d) U-Net, (e) TransUNet, (f) CMUNet, (g) CMUNeXt, (h) GGNet, (i) BGNet, (j) BFNet.

Fig. 8: Visual comparison of segmentation results of the proposed model with seven state-of-the-art methods on the TN3K dataset. (a) original image, (b) GT, (c) ours, (d) U-Net, (e) TransUNet, (f) CMUNet, (g) CMUNeXt, (h) GGNet, (i) BGNet, (j) BFNet.

Traditional Segmentation Methods
Model	Dice $\%\color[rgb]{1,0,0}{\uparrow}$	IoU $\%\color[rgb]{1,0,0}{\uparrow}$	HD95 $\color[rgb]{1,0,0}{\downarrow}$	Recall $\%\color[rgb]{1,0,0}{\uparrow}$	Acc $\%\color[rgb]{1,0,0}{\uparrow}$
U-Net (MICCAI2015) [ 23 ]	76.60	65.91	19.38	82.71	96.79
FCN-8s (CVPR2015) [ 16 ]	78.07	67.69	16.24	82.63	96.91
DeeplabV3+ (ECCV2018) [ 7 ]	79.66	69.61	14.52	84.60	96.90
AttUNet (MIDL2018) [ 20 ]	79.01	69.29	16.13	83.61	96.85
U-Net++ (TMI2020) [ 43 ]	78.85	68.99	16.55	82.47	96.87

TABLE III: Results on the TN3K dataset.

Model	Dice $\%\color[rgb]{1,0,0}{\uparrow}$	IoU $\%\color[rgb]{1,0,0}{\uparrow}$	HD95 $\color[rgb]{1,0,0}{\downarrow}$	Recall $\%\color[rgb]{1,0,0}{\uparrow}$	Acc $\%\color[rgb]{1,0,0}{\uparrow}$
Traditional Segmentation Methods
U-Net (MICCAI2015) [ 23 ]	78.13	66.91	10.44	80.76	98.71
FCN-8s (CVPR2015) [ 16 ]	79.05	67.74	8.62	81.01	98.79
DeeplabV3+ (ECCV2018) [ 7 ]	79.74	67.97	7.94	82.09	98.79
AttUNet (MIDL2018) [ 20 ]	78.76	67.52	9.84	80.68	98.80
U-Net++ (TMI2020) [ 43 ]	79.65	68.50	8.94	80.80	98.82

TABLE IV: Results on the BP dataset.

Fig. 9: Visual comparison of segmentation results of the proposed model with seven state-of-the-art methods on the BP dataset. (a) original image, (b) GT, (c) ours, (d) U-Net, (e) TransUNet, (f) CMUNet, (g) CMUNeXt, (h) GGNet, (i) BGNet, (j) BFNet.

BD	BGFE	SAAM	Dice $\%\color[rgb]{1,0,0}{\uparrow}$	IoU $\%\color[rgb]{1,0,0}{\uparrow}$	HD95 $\color[rgb]{1,0,0}{\downarrow}$	Recall $\%\color[rgb]{1,0,0}{\uparrow}$	Acc $\%\color[rgb]{1,0,0}{\uparrow}$
$\times$	$\times$	$\times$	83.05	74.91	14.34	83.80	96.95
✓	$\times$	$\times$	83.61	75.25	14.25	84.55	97.20
✓	✓	$\times$	84.95	76.71	12.91	86.79	96.85
$\times$	$\times$	✓	83.35	75.35	14.02	84.24	97.07
✓	✓	✓	85.34	77.52	11.66	87.95	97.16

TABLE V: ABLATION EXPERIMENT OF THE PROPOSED MODULES ON THE BUSI Dataset. BD means the boundary detection module, BGFE means the boundary-guided feature enhancement module, and SAAM means the adaptive aggregation module.

Fig. 10: The visual segmentation results with different modules on the BUSI dataset. (a) original image, (b) GT, (c) baseline model, (d) baseline + BD, (e) baseline + BD + BGFE, (f) baseline + SAAM, (g) baseline + BD + BGFE + SAAM.

Fig. 11: Some comparison examples of baseline model and our proposed method. (a) original image, (b) GT, (c) heat map w/o BGFE, (d) heat map with BGFE. From the above visualization results, we can find that the BGFE could effectively decrease the background noise and fucus on the lesion regions.

Fig. 12: Visualization of boundary uncertainty and the proposed boundary attention maps. (a) original image, (b) GT, (c) GT boundary, (d) prediction map of baseline model, (e) error map, (f) boundary attention map generated by our method.

Fig. 13: Different operations of boundary in BGFE. (a) refers to w/o DW

3\times 3

and

5\times 5

, (b) refers to w/. DW

3\times 3

, (c) refers to w/. DW

3\times 3

and

5\times 5

, which is used in BGFE.

	DW $3\times 3$	DW $5\times 5$	Dice $\%\color[rgb]{1,0,0}{\uparrow}$	IoU $\%\color[rgb]{1,0,0}{\uparrow}$	HD95 $\color[rgb]{1,0,0}{\downarrow}$	Recall $\%\color[rgb]{1,0,0}{\uparrow}$	Acc $\%\color[rgb]{1,0,0}{\uparrow}$
(a)	$\times$	$\times$	83.59	75.41	14.95	85.83	96.50
(b)	$\checkmark$	$\times$	84.47	76.63	13.05	85.82	96.96
(c)	$\checkmark$	$\checkmark$	85.34	77.52	11.66	87.95	97.16

TABLE VI: Results of Ablation study of different operations for boundary on the BUSI dataset. Here “DW

3\times 3

” and “DW

5\times 5

” mean the Depth-wise convolutions with kernel size as

3\times 3

and

5\times 5

	Dice $\%\color[rgb]{1,0,0}{\uparrow}$	IoU $\%\color[rgb]{1,0,0}{\uparrow}$	HD95 $\color[rgb]{1,0,0}{\downarrow}$	Recall $\%\color[rgb]{1,0,0}{\uparrow}$	Acc $\%\color[rgb]{1,0,0}{\uparrow}$
Addition	84.30	76.09	14.11	86.16	96.61
Multiply	84.62	76.71	12.97	86.89	96.63
Concatenate	84.83	76.84	12.74	87.14	96.79
BGFE (Ours)	85.34	77.52	11.66	87.95	97.16

TABLE VII: Comparison with various feature enhancement methods on the BUSI dataset.

Encoder	Decoder	Dice $\%\color[rgb]{1,0,0}{\uparrow}$	IoU $\%\color[rgb]{1,0,0}{\uparrow}$	HD95 $\color[rgb]{1,0,0}{\downarrow}$	Recall $\%\color[rgb]{1,0,0}{\uparrow}$	Acc $\%\color[rgb]{1,0,0}{\uparrow}$
$\times$	$\times$	83.35	75.35	14.02	84.24	97.07
✓	$\times$	83.53	75.63	13.97	85.09	97.10
$\times$	✓	85.34	77.52	11.66	87.95	97.16
✓	✓	84.59	76.50	12.86	86.72	97.13

TABLE VIII: Performance comparison of integrating the BGFE at different network stages on the BUSI dataset. Here we employ the boundary detection module in the “Encoder”, “Decoder”, and both stages.

	Dice $\%\color[rgb]{1,0,0}{\uparrow}$	IoU $\%\color[rgb]{1,0,0}{\uparrow}$	HD95 $\color[rgb]{1,0,0}{\downarrow}$	Recall $\%\color[rgb]{1,0,0}{\uparrow}$	Acc $\%\color[rgb]{1,0,0}{\uparrow}$
$\lambda_{2}=0.0$	83.97	75.96	13.99	86.27	96.60
$\lambda_{2}=0.1$	84.35	76.21	13.44	86.86	96.66
$\lambda_{2}=0.3$	84.48	76.33	13.02	87.37	96.52
$\lambda_{2}=0.5$	85.13	77.02	12.54	87.28	96.93
$\lambda_{2}=0.7$	85.34	77.52	11.66	87.95	97.16
$\lambda_{2}=0.9$	84.68	76.55	13.09	87.31	96.67

TABLE IX: Ablation Study of the Loss Weight

\lambda_{2}

on the BUSI Validation Dataset. We set the

\lambda_{2}

0,0.1,0.3,0.5,0.7,0.9

	Params(M) $\color[rgb]{1,0,0}{\downarrow}$	GFLOPs $\color[rgb]{1,0,0}{\downarrow}$	Dice $\%\color[rgb]{1,0,0}{\uparrow}$	IoU $\%\color[rgb]{1,0,0}{\uparrow}$	HD95 $\color[rgb]{1,0,0}{\downarrow}$
U-Net	34.53	65.52	75.57	67.38	23.61
U-Net + BGFE	42.51 $\color[rgb]{1,0,0}({\uparrow}8.01)$	90.17 $\color[rgb]{1,0,0}({\uparrow}24.65)$	84.10 $\color[rgb]{1,0,0}({\uparrow}8.53)$	76.31 $\color[rgb]{1,0,0}({\uparrow}8.93)$	13.83 $\color[rgb]{1,0,0}({\downarrow}9.78)$
TransUNet	93.23	32.23	82.94	74.79	16.37
TransUNet + BGFE	95.21 $\color[rgb]{1,0,0}({\uparrow}1.98)$	37.26 $\color[rgb]{1,0,0}({\uparrow}5.03)$	83.88 $\color[rgb]{1,0,0}({\uparrow}0.94)$	75.80 $\color[rgb]{1,0,0}({\uparrow}1.01)$	14.45 $\color[rgb]{1,0,0}({\downarrow}1.92)$
CMUNet	49.93	91.25	82.62	74.64	14.79
CMUNet + BGFE	57.91 $\color[rgb]{1,0,0}({\uparrow}7.98)$	115.87 $\color[rgb]{1,0,0}({\uparrow}24.62)$	83.93 $\color[rgb]{1,0,0}({\uparrow}1.31)$	75.81 $\color[rgb]{1,0,0}({\uparrow}1.17)$	13.23 $\color[rgb]{1,0,0}({\downarrow}1.56)$

TABLE X: Cross-architecture evaluation of the BGFE module on the BUSI dataset. The proposed module yields consistent performance improvements across all baseline models.

Methods	BUSI $\rightarrow$ BUSI		Dataset B $\rightarrow$ BUSI
Methods	Dice $\%\color[rgb]{1,0,0}{\uparrow}$	IoU $\%\color[rgb]{1,0,0}{\uparrow}$	Dice $\%\color[rgb]{1,0,0}{\uparrow}$	IoU $\%\color[rgb]{1,0,0}{\uparrow}$
U-Net (MICCAI2015) [ 23 ]	75.57	67.38	57.50	47.59
TransUNet (ARXIV2021) [ 6 ]	82.94	74.79	58.20	47.95
GGNet (MIA2021) [ 34 ]	82.63	74.23	64.26	54.86
BGNet (IJCAI2022) [ 26 ]	83.05	74.98	54.79	45.21
CMUNet (ISBI2023) [ 28 ]	82.62	74.64	56.99	47.80

TABLE XI: Robustness analysis on the BUSI and Dataset B. BUSI

\rightarrow

Dataset B means the strategy of training on the BUSI dataset and validation on the Dataset B.

Model	Pub’ Year	Parameters (M) $\color[rgb]{1,0,0}{\downarrow}$	FLOPs (G) $\color[rgb]{1,0,0}{\downarrow}$
U-Net	MICCAI2015 [ 23 ]	34.53	65.59
Att U-Net	MIDL2018 [ 20 ]	34.88	66.70
GGNet	MIA2021 [ 34 ]	59.47	18.40
TransUnet	Arxiv2021 [ 6 ]	93.20	24.70
BGNet	IJCAI2022 [ 26 ]	77.80	22.16
AAU-Net	TMI2023 [ 4 ]	158.51	249.37

TABLE XII: The results of parameters and FLOPs comparison for different methods. We employ the “Parameters” and “FLOPS” to evaluate the complexity of different methods.

Fig. 14: Some failure cases of PBE-UNet on the BUSI dataset. (a) original image, (b) GT, (c) ours.

Ultrasound Image Segmentation

Ultrasound image segmentation delineates organs, lesions, and other structures in ultrasound scans for measurement, diagnosis, and image-guided care.

See all 73 papers

Risk-Routed Implicit Boundary Refinement for Robust Ultrasound Image Segmentation
Jul 23, 2026Jingguo Qu, Xinyang Han, Xiang Wang +8Ultrasound Image Segmentation Implicit Neural Representations
Department of Health Technology and Informatics, The Hong Kong Polytechnic University, Hong Kong, China · Centre for Smart Health and School of Nursing, The Hong Kong Polytechnic University, Hong Kong, China · Department of Imaging and Interventional Radiology, The Chinese University of Hong Kong, Hong Kong, China
An Edge-aware Prompt-enhanced SAM for Ultrasound Image Segmentation
Jul 8, 2026Wenhao Li, Fangyi Liu, Bo DuUltrasound Image Segmentation Promptable Image Segmentation
Key Laboratory of Aerospace Information Security and Trusted Computing, Ministry of Education, School of Cyber Science and Engineering, Wuhan University, Wuhan, China
MoWE: Spatially Adaptive Mixture-of-Wavelet-Experts Pooling for Ultrasound Lesion Segmentation
Oct 18, 2026Reaz Hassan Joarder, Sidratul Tanzila Tasmi, Alif Arshad Bakshi +2Ultrasound Image Segmentation Mixture of Experts
Islamic University of Technology, Bangladesh · United International University, Bangladesh

Jul 23, 2026J/K move · Enter open · S save

Risk-Routed Implicit Boundary Refinement for Robust Ultrasound Image Segmentation

Jingguo Qu, Xinyang Han, Xiang Wang, Yuqi Yang+7

Department of Health Technology and Informatics, The Hong Kong Polytechnic University, Hong Kong, China · Centre for Smart Health and School of Nursing, The Hong Kong Polytechnic University, Hong Kong, China · Department of Imaging and Interventional Radiology, The Chinese University of Hong Kong, Hong Kong, China

Ultrasound Image Segmentation Implicit Neural Representations Medical Image Segmentation Boundary-Aware Segmentation

Open paper AskPDF arXiv

Accurate boundary delineation in medical ultrasound (US) images remains challenging because of speckle noise, low contrast, acoustic shadowing, and acquisition variability. We propose Risk-routed Implicit Boundary Refinement (RIBR), a compact framework that refines convolutional segmentation through spatially weighted implicit residuals. A coordinate-conditioned implicit branch uses regional features to predict logit corrections, and a learned gate controls their contribution using predictive uncertainty and local appearance. Geometry- and speckle-aware regularization provides contour supervision and discourages probability fluctuations away from boundaries. We compare RIBR with ten baselines on nine datasets covering lymph nodes, breast lesions, thyroid nodules, and prostate, including four external lymph-node and breast cohorts. Across three seeds, RIBR achieves the best overall dataset macro-average among the evaluated configurations, with

76.67{\pm}0.42\%

Dice and a 95th-percentile Hausdorff distance (HD95) of

29.55{\pm}0.55

pixels on the resized evaluation grid. It obtains the lowest HD95 on all nine datasets, with task-dependent Dice rankings. These results support controlled implicit residual refinement for boundary-sensitive US segmentation. Source code is available at https://github.com/jinggqu/ribr.

Abstract

Ultrasound Image Segmentation

Explore similar work

Risk-Routed Implicit Boundary Refinement for Robust Ultrasound Image Segmentation

An Edge-aware Prompt-enhanced SAM for Ultrasound Image Segmentation

MoWE: Spatially Adaptive Mixture-of-Wavelet-Experts Pooling for Ultrasound Lesion Segmentation