基于逐通道空间自适应选择核卷积与双向边界感知机制的乳腺超声图像病变分割网络

doi:10.12141/j.issn.1000-565X.250230

摘要/Abstract

摘要：

乳腺癌是全球女性最常见的恶性肿瘤之一，准确的病变分割对于乳腺癌的早期诊断与治疗具有重要意义。然而，由于病变形态的多样性以及超声成像机制的复杂性，现有基于深度学习的乳腺超声图像病变分割方法在分割准确性方面仍面临巨大挑战。为进一步提升乳腺超声图像中病变区域的分割精度，该文基于经典U-Net架构，提出了一种新型乳腺超声图像病变分割网络（CWSASKM-BBAM-Net）。首先，在网络中引入逐通道空间自适应选择核卷积模块（CWSASKM），根据不同通道的语义特征为每个空间位置自适应选择感受野大小，以增强多尺度信息的建模能力；然后，引入双向边界感知机制（BBAM），通过融合正向与反向注意力，对目标显著区域及其边界进行协同建模，同时逐步提升对非显著区域与病变区域的区分能力，以进一步强化边界信息的表达；最后，在3组公开乳腺超声图像数据集（BUSI、UDIAT和STU）上开展分割实验。结果表明：该方法在数据集BUSI上的杰卡德指数、精确率、召回率和Dice相似系数分别为71.97%、82.85%、81.40%和80.44%，较次优方法分别提升1.69、1.05、1.28和1.84个百分点；在数据集UDIAT上，这4项指标分别达到78.14%、88.31%、86.73%和86.10%，较次优方法分别提升了2.75、2.04、0.56和2.01个百分点；在外部数据集STU上，该方法也取得了优于其他方法的整体表现。实验结果表明，CWSASKM-BBAM-Net在乳腺超声图像分割任务中展现出更优的整体性能。

关键词: 乳腺超声图像, 病变分割, 自适应选择核卷积, 双向边界感知机制

Abstract:

Breast cancer is the most common malignancy among women worldwide, and accurate lesion segmentation is of great importance for its early diagnosis and treatment. However, due to the high morphological variability of lesions and the inherent complexity of ultrasound imaging, existing deep-learning-based methods still face significant challenges in achieving satisfactory segmentation accuracy for breast ultrasound images. To address this limitation, this study proposed a novel breast-lesion segmentation network, termed CWSASKM-BBAM-Net, which is built upon the classical U-Net architecture. First, a Channel-Wise Spatially Adaptive Selective Kernel Convolution Module (CWSASKM) was introduced to adaptively adjust the receptive field size for each spatial location based on channel-specific semantic features, thereby enhancing the network’s multi-scale representation learning. Second, a Bidirectional Boundary-Aware Mechanism (BBAM) was designed to jointly model salient regions and their boundaries by integrating forward and reverse attention. This mechanism progressively improves the discrimination between non-salient areas and lesion areas, thereby refining boundary delineation. Extensive experiments were conducted on three public breast ultrasound datasets (BUSI, UDIAT, and STU). On BUSI dataset, the proposed method achieved Jaccard index, precision, recall, and Dice similarity coefficient of 71.97%, 82.85%, 81.40%, and 80.44%, respectively, outperforming the second-best approach by margins of 1.69, 1.05, 1.28, and 1.84 percentage points. On UDIAT, it attained 78.14%, 88.31%, 86.73%, and 86.10% on these metrics, with improvements of 2.75, 2.04, 0.56, and 2.01 percentage points, respectively. The proposed method also demonstrated superior overall performance on the external STU dataset. These results collectively demonstrate that CWSASKM-BBAM-Net achieves state-of-the-art performance in breast ultrasound image segmentation tasks.

Key words: breast ultrasound image, lesion segmentation, adaptive selective kernel convolution, bidirectional boundary-aware mechanism

中图分类号:

TP391.41

王洁, 李璐瑶. 基于逐通道空间自适应选择核卷积与双向边界感知机制的乳腺超声图像病变分割网络[J]. 华南理工大学学报(自然科学版), 2026, 54(2): 77-90.

WANG Jie, LI Luyao. A Breast Ultrasound Images Lesion Segmentation Network Based on Channel-Wise Spatially Adaptive Selective Kernel Convolution and Bidirectional Boundary-Aware Mechanism[J]. Journal of South China University of Technology(Natural Science Edition), 2026, 54(2): 77-90.

图/表 16

图１

图2

图3

图4

图5

图6

图7

表1

表2

表3

表4

表5

表6

表7

表8

表9

参考文献 22

[1]	BRAY F， LAVERSANNE M， SUNG H，et al ．Global cancer statistics 2022：GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries［J］．CA：A Cancer Journal for Clinicians，2024，74（3）：229-263.
[2]	KIM J， HARPER A， McCORMACK V，et al ．Global patterns and trends in breast cancer incidence and mortality across 185 countries［J］．Nature Medicine，2025，31（4）：1154-1162.
[3]	FERLAY J， COLOMBET M， SOERJOMATARAM I，et al ．Estimating the global cancer incidence and mortality in 2018：GLOBOCAN sources and methods［J］．International Journal of Cancer，2019，144（8）：1941-1953.
[4]	GHONCHEH M， POURNAMDAR Z， SALEHINIYA H ．Incidence and mortality and epidemiology of breast cancer in the world［J］．Asian Pacific Journal of Cancer Prevention，2016，17（S3）：43-46.
[5]	JIN J ．Breast cancer screening：benefits and harms［J］．JAMA，2014，312（23）：2585.
[6]	MARMOT M G， ALTMAN D G， CAMERON D A，et al ．The benefits and harms of breast cancer screening：an independent review［J］．British Journal of Cancer，2013，108：2205-2240.
[7]	BERG W A， BANDOS A I， MENDELSON E B，et al ．Ultrasound as the primary screening test for breast cancer：analysis from ACRIN 6666［J］．Journal of the National Cancer Institute，2016，108（4）：djv367/1-8.
[8]	JOO S， YANG Y S， MOON W K，et al ．Computer-aided diagnosis of solid breast nodules：use of an artificial neural network based on multiple sonographic features［J］．IEEE Transactions on Medical Imaging，2004，23（10）：1292-1300.
[9]	MOON W K， LO C M， CHEN R T，et al ．Tumor detection in automated breast ultrasound images using quantitative tissue clustering［J］．Medical Physics，2014，41（4）：042901/1-9.
[10]	RONNEBERGER O， FISCHER P， BROX T ．U-Net：convolutional networks for biomedical image segmentation［C］∥ Proceedings of the 18th International Conference on Medical Image Computing and Computer-Assisted Intervention．Munich：Springer，2015：234-241.
[11]	ZHOU Z， SIDDIQUEE M M R， TAJBAKHSH N，et al ．UNet++：redesigning skip connections to exploit multiscale features in image segmentation［J］．IEEE Transactions on Medical Imaging，2020，39：1856-1867.
[12]	BYRA M， JAROSIK P， SZUBERT A，et al ．Breast mass segmentation in ultrasound with selective kernel U-Net convolutional neural network［J］．Biomedical Signal Processing and Control，2020，61：102027/1-10.
[13]	CHEN G， LI L， DAI Y，et al ．AAU-Net：an adaptive attention U-Net for breast lesions segmentation in ultrasound images［J］．IEEE Transactions on Medical Imaging，2023，42：1289-1300.
[14]	CHEN G， ZHOU L， ZHANG J，et al ．ESKNet：an enhanced adaptive selection kernel convolution for ultrasound breast tumors segmentation［J］．Expert Systems with Applications，2024，246：123265/1-17.
[15]	ABRAHAM N， KHAN N M ．A novel focal Tversky loss function with improved attention U-Net for lesion segmentation［C］∥ Proceedings of IEEE the 16th International Symposium on Biomedical Imaging．Venice：IEEE，2019：683-687.
[16]	OKTAY O， SCHLEMPER J， FOLGOC L L，et al ．Attention U-Net：learning where to look for the pancreas［C］∥ Proceedings of the 1st Conference on Medical Imaging with Deep Learning.Amsterdam：OpenReview，2018：1-10.
[17]	YAN Y， LIU Y， WU Y，et al ．Accurate segmentation of breast tumors using AE U-Net with HDC model in ultrasound images［J］．Biomedical Signal Processing and Control，2022，72：103299/1-7.
[18]	AL-DHABYANI W， GOMAA M， KHALED H，et al ．Dataset of breast ultrasound images［J］．Data in Brief，2020，28：104863/1-5.
[19]	YAP M H， GOYAL M， OSMAN F，et al ．Breast ultrasound region of interest detection and lesion localisation［J］．Artificial Intelligence in Medicine，2020，107：101880/1-8.
[20]	ZHUANG Z， LI N， JOSEPH RAJ A N，et al ．An RDAU-NET model for lesion segmentation in breast ultrasound images［J］．PLoS One，2019，14（8）：e0221535/1-23.
[21]	SHAREEF B， XIAN M， VAKANSKI A ．STAN：small tumor-aware network for breast ultrasound image segmentation［C］∥ Proceedings of IEEE the 17th International Symposium on Biomedical Imaging．Iowa City：IEEE，2020：1-5.
[22]	CUI K， TIAN Q， WANG H，et al ．An improved framework for breast ultrasound image segmentation with multiple branches depth perception and layer compression residual module［J］．Engineering Applications of Artificial Intelligence，2025，146：110265/1-13.

数据集	通道数	参数量/10⁶	均值±标准差
数据集	通道数	参数量/10⁶	J/%	P/%	R/%	c_Dice/%
BUSI	8	14.60	71.66 ± 2.10	82.74 ± 2.67	80.89 ± 2.30	80.08 ± 2.11
	16	14.63	71.97 ± 1.84	82.85 ± 1.05	81.40 ± 2.22	80.44 ± 1.73
	32	14.73	71.85 ± 1.92	83.12 ± 1.68	81.25 ± 2.45	80.52 ± 1.88
UDIAT	8	14.60	77.82 ± 2.45	87.95 ± 4.12	86.45 ± 1.78	85.88 ± 2.19
	16	14.63	78.14 ± 2.39	88.31 ± 3.99	86.73 ± 1.62	86.10 ± 2.13
	32	14.73	78.35 ± 2.85	88.55 ± 4.35	86.85 ± 2.10	86.35 ± 2.55

方法	均值±标准差
方法	J/%	P/%	R/%	c_Dice/%
U-Net	61.31 ± 1.74	75.91 ± 1.85	73.84 ± 1.87	71.57 ± 1.85
UNet++	64.66 ± 1.39	75.65 ± 2.58	78.15 ± 2.17	73.67 ± 1.62
Attention U-Net	64.87 ± 1.89	76.88 ± 1.93	77.39 ± 4.32	73.82 ± 2.22
MADU-Net	64.69 ± 1.34	78.72 ± 1.10	75.88 ± 1.82	74.29 ± 1.39
STAN	65.72 ± 2.89	80.27 ± 2.59	76.13 ± 2.94	74.96 ± 2.88
SK U-Net	69.08 ± 1.99	80.88 ± 1.58	78.41 ± 1.92	77.44 ± 1.77
AE U-Net	66.81 ± 1.77	80.44 ± 2.13	76.45 ± 2.37	75.64 ± 1.86
AAU-Net	69.35 ± 2.60	80.17 ± 2.45	79.87 ± 2.08	77.91 ± 2.25
ESKNet	70.28 ± 1.96	81.80 ± 1.18	80.04 ± 1.35	78.60 ± 1.79
MBDPNet	69.19 ± 2.38	80.41 ± 0.98	80.12 ± 2.38	78.21 ± 2.09
CWSASKM-BBAM-Net	71.97 ± 1.84	82.85 ± 1.05	81.40 ± 2.22	80.44 ± 1.73

方法	均值±标准差
方法	J/%	P/%	R/%	c_Dice/%
U-Net	70.30 ± 3.37	81.53 ± 2.34	81.86 ± 3.38	79.43 ± 2.71
UNet++	74.30 ± 4.17	86.27 ± 3.63	82.72 ± 4.40	82.37 ± 3.59
Attention U-Net	73.69 ± 3.78	84.93 ± 3.95	83.25 ± 2.25	81.93 ± 2.76
MADU-Net	71.44 ± 5.91	84.03 ± 2.56	80.00 ± 5.96	79.86 ± 4.55
STAN	67.76 ± 4.09	81.77 ± 3.85	77.75 ± 4.45	76.49 ± 3.65
SK U-Net	74.54 ± 3.75	84.51 ± 5.69	84.32 ± 1.84	82.69 ± 3.75
AE U-Net	72.43 ± 3.33	84.22 ± 2.74	82.91 ± 3.65	80.90 ± 2.62
AAU-Net	73.76 ± 2.81	84.82 ± 3.37	83.01 ± 2.93	82.20 ± 2.72
ESKNet	75.39 ± 2.39	85.33 ± 3.06	86.17 ± 2.01	84.09 ± 2.33
MBDPNet	75.16 ± 2.54	84.82 ± 3.27	86.09 ± 1.65	83.81 ± 2.02
CWSASKM-BBAM-Net	78.14 ± 2.39	88.31 ± 3.99	86.73 ± 1.62	86.10 ± 2.13

方法	均值±标准差
方法	J/%	P/%	R/%	c_Dice/%
U-Net	73.57 ± 1.25	85.31 ± 2.49	83.80 ± 2.36	82.50 ± 1.45
UNet++	74.75 ± 0.93	84.42 ± 2.73	87.46 ± 1.40	83.97 ± 0.90
Attention U-Net	72.30 ± 1.29	84.77 ± 2.50	82.72 ± 2.56	81.51 ± 1.55
MADU-Net	72.65 ± 0.84	85.25 ± 1.40	83.30 ± 1.54	82.59 ± 0.69
STAN	74.68 ± 0.63	84.09 ± 0.84	86.73 ± 1.40	84.31 ± 0.91
SK U-Net	77.41 ± 0.39	84.08 ± 0.28	90.24 ± 0.37	86.13 ± 0.30
AE U-Net	73.78 ± 0.64	83.07 ± 0.92	86.60 ± 1.43	83.65 ± 0.97
AAU-Net	76.75 ± 1.00	83.49 ± 1.01	91.49 ± 0.58	86.21 ± 0.76
ESKNet	77.60 ± 1.01	84.20 ± 1.09	91.71 ± 0.56	86.78 ± 0.75
MBDPNet	78.16 ± 0.51	84.79 ± 0.35	90.46 ± 0.34	86.67 ± 0.34
CWSASKM-BBAM-Net	78.47 ± 0.90	84.84 ± 1.30	92.36 ± 0.85	87.40 ± 0.64

方法	均值±标准差
方法	J/%	P/%	R/%	c_Dice/%
U-Net	70.28 ± 0.83	91.98 ± 0.67	75.80 ± 1.27	80.32 ± 0.82
UNet++	72.45 ± 2.69	93.23 ± 1.39	76.99 ± 3.06	81.77 ± 2.55
Attention U-Net	70.26 ± 2.50	91.51 ± 0.84	75.38 ± 3.53	79.87 ± 2.22
MADU-Net	68.36 ± 1.84	90.75 ± 1.04	73.56 ± 2.61	78.16 ± 1.90
STAN	73.74 ± 0.89	91.78 ± 0.92	79.93 ± 0.68	83.35 ± 0.58
SK U-Net	72.67 ± 1.99	92.76 ± 0.97	77.58 ± 2.31	82.44 ± 1.57
AE U-Net	71.80 ± 2.05	92.85 ± 0.75	76.58 ± 2.37	81.75 ± 1.63
AAU-Net	77.36 ± 1.49	93.41 ± 0.34	82.39 ± 1.49	86.28 ± 1.09
ESKNet	75.81 ± 1.94	93.83 ± 0.24	80.52 ± 2.17	84.58 ± 1.75
MBDPNet	76.88 ± 1.28	93.06 ± 0.40	82.20 ± 1.18	85.96 ± 1.03
CWSASKM-BBAM-Net	78.55 ± 1.05	94.16 ± 0.45	83.25 ± 0.88	87.04 ± 0.85