SG-UNet：基于全局注意力和自校准卷积增强的黑色素瘤分割模型

doi:10.12122/j.issn.1673-4254.2025.06.21

南方医科大学学报 ›› 2025, Vol. 45 ›› Issue (6): 1317-1326.doi: 10.12122/j.issn.1673-4254.2025.06.21

SG-UNet：基于全局注意力和自校准卷积增强的黑色素瘤分割模型

计寰宇¹(), 王蕊², 高盛祥¹^,⁴, 车文刚¹^,³()

^1.昆明理工大学，信息工程与自动化学院，云南昆明 650500
^2.昆明城市学院数据科学与工程学院，云南昆明 650106
^3.昆明理工大学，云南省计算机技术应用重点实验室，云南昆明 650500
^4.昆明理工大学，云南省人工智能重点实验室，云南昆明 650500

收稿日期:2024-11-13 出版日期:2025-06-20 发布日期:2025-06-27
通讯作者: 车文刚 E-mail:2285873874@qq.com;goooglethink@gmail.com
作者简介:计寰宇，在读硕士研究生，E-mail: 2285873874@qq.com
基金资助:
国家自然科学基金(U23A20388);云南省重点研发计划(202303AP140008);云南省基础研究项目(202301AT070393);昆明理工大学“双一流”科技专项(202402AG050007)

SG-UNet: a melanoma segmentation model enhanced with global attention and self-calibrated convolution

Huanyu JI¹(), Rui WANG², Shengxiang GAO¹^,⁴, Wengang CHE¹^,³()

^1.Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming 650500, China
^2.School of Data Science and Engineering, Kunming City College, Kunming 650106, China
^3.Yunnan Key Laboratory of Computer Technology Applications, Kunming University of Science and Technology, Kunming 650500, China
^4.Yunnan Key Laboratory of Artificial Intelligence, Kunming University of Science and Technology, Kunming 650500, China

Received:2024-11-13 Online:2025-06-20 Published:2025-06-27
Contact: Wengang CHE E-mail:2285873874@qq.com;goooglethink@gmail.com
Supported by:
National Natural Science Foundation of China(U23A20388)

摘要/Abstract

摘要：

目的提出了一种新的黑色素瘤分割模型SG-UNet，以提高黑色素瘤皮肤镜图像的精确分割。通过分割后边界特征评估，可以更准确地识别诊断黑色素瘤从而辅助早期诊断。方法使用一种U形结构的卷积神经网络UNet，对其主干、跳跃连接和下采样池化部分进行改进。在主干部分，我们将UNet的下采样部分参考Vgg的结构将卷积数量由10个增加到13个加深网络层次来捕获更加精细的特征表示。为了进一步提升特征提取和细节识别的能力，主干部分将传统的卷积替换为自校准卷积增强模型对空间维度和通道维度特征的捕获能力。同时，在池化部分将哈尔小波下采样替换原有的池化层实现更有效的多尺度特征融合，并降低特征图的空间分辨率。接着将全局注意力机制融入到每一层的跳跃连接中更好地理解图像的上下文信息。结果实验结果表明SG-UNet在ISIC 2017和ISIC 2018数据集上的分割效果对比目前其他先进分割模型得到明显提升。在ISIC 2017和ISIC 2018数据集上Dice，IoU分别达到了92.41%，86.62%和92.31%，86.48%。结论实验结果证实，所提出的方法能够有效实现黑色素瘤的精确分割。

关键词: 图像分割, 全局注意力机制, 黑色素瘤, UNet, 自校准卷积, 哈尔小波下采样, SG-UNet

Abstract:

Objective We propose a new melanoma segmentation model, SG-UNet, to enhance the precision of melanoma segmentation in dermascopy images to facilitate early melanoma detection. Methods We utilized a U-shaped convolutional neural network, UNet, and made improvements to its backbone, skip connections, and downsampling pooling sections. In the backbone, with reference to the structure of VGG, we increased the number of convolutions from 10 to 13 in the downsampling part of UNet to achieve a deepened network hierarchy that allowed capture of more refined feature representations. To further enhance feature extraction and detail recognition, we replaced the traditional convolution the backbone section with self-calibrated convolution to enhance the model's ability to capture both spatial and channel dimensional features. In the pooling part, the original pooling layer was replaced by Haar wavelet downsampling to achieve more effective multi-scale feature fusion and reduce the spatial resolution of the feature map. The global attention mechanism was then incorporated into the skip connections at each layer to enhance the understanding of contextual information of the image. Results The experimental results showed that the SG-UNet model achieved significantly improved segmentation accuracy on ISIC 2017 and ISIC 2018 datasets as compared with other current state-of-the-art segmentation models, with Dice reached 92.41% and 86.62% and IoU reaching 92.31% and 86.48% on the two datasets, respectively. Conclusion The proposed model is capable of effective and accurate segmentation of melanoma from dermoscopy images.

Key words: image segmentation, global attention mechanism, melanoma, UNet, self-calibrated convolution, Haar wavelet downsampling, SG-UNet

计寰宇, 王蕊, 高盛祥, 车文刚. SG-UNet：基于全局注意力和自校准卷积增强的黑色素瘤分割模型[J]. 南方医科大学学报, 2025, 45(6): 1317-1326.

Huanyu JI, Rui WANG, Shengxiang GAO, Wengang CHE. SG-UNet: a melanoma segmentation model enhanced with global attention and self-calibrated convolution[J]. Journal of Southern Medical University, 2025, 45(6): 1317-1326.

图/表 14

图1 SG-UNet模型结构图

Fig.1 SG-UNet model structure diagram.

图2 GAM注意力机制结构图

Fig.2 GAM structure diagram.

图3 皮肤镜下图像与GAM注意力机制生成的热力图

Fig.3 Heat map generated by dermoscopic images with global attention mechanism (GAM).

图4 SCConv结构图

Fig.4 SCConv structure diagram.

图5 HWD总体结构

Fig.5 Framework structure of Haar wavelet downsampling (HWD).

图6 Haar小波变换

Fig.6 Haar wavelet transform.

图7 黑色素瘤经Haar小波变换后得到的4个分量

Fig.7 Four components obtained from melanoma after Haar wavelet transform. A: Low-frequency information; H: Horizontal high-frequency information; V: Vertical high-frequency information; D: Diagonal high-frequency information.

图8 SG-UNet在ISIC2018(A)和ISIC 2017(B)上实验结果

Fig.8 Experimental results of SG-UNet on ISIC 2018(A) and ISIC 2017(B) datasets.

表1 ISIC 2018数据集各模型性能指标

Tab.1 Performance metrics for each model in the ISIC 2018 dataset

Model	Dice	IoU	Recall	Precision
UNet^[4]	89.46	82.13	90.44	89.64
UNet++^[25]	89.65	82.36	89.98	90.30
AttUNet^[26]	90.03	82.90	89.73	91.27
ResUNet^[27]	90.86	84.26	91.51	91.20
ResUNet++^[28]	91.29	84.87	91.94	91.45
FAT-Net^[7]	89.03	82.02	91.00	-
CA-net^[29]	85.35	74.44	78.76	93.14
CKDNet^[30]	87.79	80.41	90.55	-
PraNet^[34]	87.37	77.57	86.64	88.11
MSCNet^[36]	90.51	82.67	89.61	91.84
DAGAN^[33]	88.07	81.13	90.72	-
CPFNet^[31]	87.69	79.88	89.53	-
Ours	92.34	86.48	93.05	92.21

表2 ISIC 2017数据集各模型性能指标

Tab.2 Performance metrics for each model in the ISIC 2017 dataset

Model	Dice	IoU	Recall	Precision
UNet^[4]	89.20	81.98	91.24	88.81
UNet++^[25]	89.50	82.23	91.22	88.92
AttUNet^[26]	90.67	83.98	92.07	90.22
ResUNet^[27]	91.98	85.99	93.21	91.51
ResUNet++^[28]	90.60	84.05	92.62	89.96
FAT-Net^[7]	85.00	76.53	83.92	-
CA-net^[29]	86.30	75.91	85.45	87.18
DAGAN^[33]	84.25	75.94	83.63	-
SESV^[35]	83.92	75.31	83.26	-
PraNet^[34]	87.71	78.23	87.58	87.84
MB-DCNN^[32]	84.27	76.03	83.25	-
MSCNet^[36]	90.99	83.47	91.11	90.87
CPFNet^[31]	84.03	75.46	83.44	-
Ours	92.41	86.62	93.70	91.80

表3 ISIC2017数据集上的消融实验结果

Tab.3 Ablation study results on the ISIC dataset

Model	SCConv	HWD	GAM	Dice	IoU	Recall	Precision
Baseline	×	×	×	89.20	81.98	91.24	88.81
Model①	×	×	×	89.26	81.78	91.96	87.87
Model②	×	√	×	91.05	84.54	92.21	90.73
Model③	×	√	√	91.39	85.05	92.94	90.75
Model④	√	√	√	92.41	86.62	93.70	91.80

表4 ISIC2018数据集上的消融实验结果

Tab.4 Ablation study results on the ISIC 2018 dataset

Model	SCConv	HWD	GAM	Dice	IoU	Recall	Precision
Baseline	×	×	×	89.46	82.13	90.44	89.64
Model①	×	×	×	90.17	83.06	91.08	90.03
Model②	×	√	×	91.35	84.89	91.84	91.54
Model③	×	√	√	91.38	84.88	91.70	91.68
Model④	√	√	√	92.34	86.48	93.05	92.21

图9 各模型在ISIC 2017数据集上的分割结果

Fig.9 Segmentation results for each model on the ISIC 2017 dataset.

图10 各模型在ISIC 2018数据集上的分割结果

Fig.10 Segmentation results for each model on the ISIC 2018 dataset.

参考文献 36

1	Mateen M, Hayat S, Arshad F, et al. Hybrid deep learning framework for melanoma diagnosis using dermoscopic medical images[J]. Diagnostics, 2024, 14(19): 2242. doi：10.3390/diagnostics14192242
2	Ali AR, Li JP, O'Shea SJ, et al. A deep learning based approach to skin lesion border extraction with a novel edge detector in dermoscopy images[C]//2019 International Joint Conference on Neural Networks (IJCNN). July 14-19, 2019, Budapest, Hungary. IEEE, 2019: 1-7. doi：10.1109/ijcnn.2019.8852134
3	Ali AR, Li JP, Yang G, et al. A machine learning approach to automatic detection of irregularity in skin lesion border using dermoscopic images[J]. PeerJ Comput Sci, 2020, 6: e268. doi：10.7717/peerj-cs.268
4	Ronneberger O, Fischer P, Brox T. U-Net: convolutional networks for biomedical image segmentation[M]//Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015. Cham: Springer International Publishing, 2015: 234-41. doi：10.1007/978-3-319-24574-4_28
5	Long J, Shelhamer E, Darrell T. Fully convolutional networks for semantic segmentation[C]//2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). June 7-12, 2015, Boston, MA, USA. IEEE, 2015: 3431-40. doi：10.1109/cvpr.2015.7298965
6	Wang G, Ma QS, Li YY, et al. A skin lesion segmentation network with edge and body fusion[J]. Appl Soft Comput, 2025, 170: 112683. doi：10.1016/j.asoc.2024.112683
7	Wu HS, Chen SH, Chen GL, et al. FAT-Net: Feature adaptive transformers for automated skin lesion segmentation[J]. Med Image Anal, 2022, 76: 102327. doi：10.1016/j.media.2021.102327
8	王娜, 贾伟, 赵雪芬, 等. 基于边缘关键点和边缘注意力的黑色素瘤图像分割方法[J].中国医学物理学杂志 [J]. 2024, 41(10): 1225-36.
9	Liu Y, Shao Z, Hoffmann N. Global attention mechanism: Retain information to enhance channel-spatial interactions [J]. 2021, DOI:10.48550/arXiv.2112.05561 .
10	Liu JJ, Hou QB, Cheng MM, et al. Improving convolutional networks with self-calibrated convolutions[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). June 13-19, 2020, Seattle, WA, USA. IEEE, 2020: 10093-102. doi：10.1109/cvpr42600.2020.01011
11	Xu GP, Liao WT, Zhang X, et al. Haar wavelet downsampling: a simple but effective downsampling module for semantic segmentation[J]. Pattern Recognit, 2023, 143: 109819. doi：10.1016/j.patcog.2023.109819
12	Niu ZY, Zhong GQ, Yu H. A review on the attention mechanism of deep learning[J]. Neurocomputing, 2021, 452: 48-62. doi：10.1016/j.neucom.2021.03.091
13	Upadhyay AK, Bhandari AK. Advances in deep learning models for resolving medical image segmentation data scarcity problem: a topical review[J]. Arch Comput Meth Eng, 2024, 31(3): 1701-19. doi：10.1007/s11831-023-10028-9
14	Woo S, Park J, Lee JY, et al. CBAM: convolutional block attention module[M]//Computer Vision-ECCV 2018. Cham: Springer International Publishing, 2018: 3-19. doi：10.1007/978-3-030-01234-2_1
15	Ioannou Y, Robertson D, Cipolla R, et al. Deep roots: improving CNN efficiency with hierarchical filter groups[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). July 21-26, 2017, Honolulu, HI, USA. IEEE, 2017: 5977-86. doi：10.1109/cvpr.2017.633
16	Porwik P, Lisowska A. The haar–wavelet transform in digital image processing: its status and achievements[J]. Instit Comp Sci, 2004,13:79-98.
17	Makandar A, Halalli B. Image enhancement techniques using highpass and lowpass filters[J]. Int J Comput Appl, 2015, 109(14): 21-7. doi：10.5120/19256-0999
18	Zaynidinov H. Digital image processing with two-dimensional haar wavelets[J]. Int J Adv Trends Comput Sci Eng, 2020, 9(3): 2729-34. doi：10.30534/ijatcse/2020/38932020
19	Codella NCF, Gutman D, Celebi ME, et al. Skin lesion analysis toward melanoma detection: a challenge at the 2017 International symposium on biomedical imaging (ISBI), hosted by the international skin imaging collaboration (ISIC)[C]//2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018). April 4-7, 2018, Washington, DC, USA. IEEE, 2018: 168-72. doi：10.1109/isbi.2018.8363547
20	Codella N, Rotemberg V, Tschandl P, et al. Skin lesion analysis toward melanoma detection 2018: a challenge hosted by the international skin imaging collaboration (ISIC)[EB/OL]. 2019: 1902.03368. . doi：10.48550/arXiv.1902.03368
21	Jiang HY, Diao ZS, Shi TY, et al. A review of deep learning-based multiple-lesion recognition from medical images: classification, detection and segmentation[J]. Comput Biol Med, 2023, 157: 106726. doi：10.1016/j.compbiomed.2023.106726
22	Soomro TA, Afifi AJ, Gao JB, et al. Strided U-Net model: retinal vessels segmentation using dice loss[C]//2018 Digital Image Computing: Techniques and Applications (DICTA). December 10-13, 2018, Canberra, ACT, Australia. IEEE, 2018: 1-8. doi：10.1109/dicta.2018.8615770
23	Cheng BW, Girshick R, Dollár P, et al. Boundary IoU: improving object-centric image segmentation evaluation[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). June 20-25, 2021, Nashville, TN, USA. IEEE, 2021: 15329-37. doi：10.1109/cvpr46437.2021.01508
24	Shamir RR, Duchin Y, Kim J, et al. Continuous dice coefficient: a method for evaluating probabilistic segmentations [J]. 2018, DOI:10.1101/306977 .
25	Zhou ZW, Siddiquee MMR, Tajbakhsh N, et al. UNet++: redesigning skip connections to exploit multiscale features in image segmentation[J]. IEEE Trans Med Imaging, 2020, 39(6): 1856-67. doi：10.1109/tmi.2019.2959609
26	Wang SH, Li L, Zhuang XH. AttU-NET: attention U-Net for brain tumor segmentation[M]//Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries. Cham: Springer International Publishing, 2022: 302-11. doi：10.1007/978-3-031-09002-8_27
27	Rahman H, Bukht TFN, Imran A, et al. A deep learning approach for liver and tumor segmentation in CT images using ResUNet[J]. Bioengineering, 2022, 9(8): 368. doi：10.3390/bioengineering9080368
28	Jha D, Smedsrud PH, Riegler MA, et al. ResUNet: an advanced architecture for medical image segmentation[C]//2019 IEEE International Symposium on Multimedia (ISM). December 9-11, 2019, San Diego, CA, USA. IEEE, 2019: 225-2255. doi：10.1109/ism46123.2019.00049
29	Gu R, Wang GT, Song T, et al. CA-net: comprehensive attention convolutional neural networks for explainable medical image segmentation[J]. IEEE Trans Med Imaging, 2021, 40(2): 699-711. doi：10.1109/tmi.2020.3035253
30	Jin QG, Cui H, Sun CM, et al. Cascade knowledge diffusion network for skin lesion diagnosis and segmentation[J]. Appl Soft Comput, 2021, 99: 106881. doi：10.1016/j.asoc.2020.106881
31	Feng SL, Zhao HM, Shi F, et al. CPFNet: context pyramid fusion network for medical image segmentation[J]. IEEE Trans Med Imaging, 2020, 39(10): 3008-18. doi：10.1109/tmi.2020.2983721
32	Xie YT, Zhang JP, Xia Y, et al. A mutual bootstrapping model for automated skin lesion segmentation and classification[J]. IEEE Trans Med Imaging, 2020, 39(7): 2482-93. doi：10.1109/tmi.2020.2972964
33	Yang G, Yu SM, Dong H, et al. DAGAN: deep de-aliasing generative adversarial networks for fast compressed sensing MRI reconstruction[J]. IEEE Trans Med Imaging, 2018, 37(6): 1310-21. doi：10.1109/tmi.2017.2785879
34	Fan DP, Ji GP, Zhou T, et al. PraNet: parallel reverse attention network for polyp segmentation[M]//Medical Image Computing and Computer Assisted Intervention-MICCAI 2020. Cham: Springer International Publishing, 2020: 263-73. doi：10.1007/978-3-030-59725-2_26
35	Xie YT, Zhang JP, Lu H, et al. SESV: accurate medical image segmentation by predicting and correcting errors[J]. IEEE Trans Med Imag, 2021, 40(1): 286-96. doi：10.1109/tmi.2020.3025308
36	Li Z, Zhang L. Multi-scale context fusion network for melanoma segmentation[J]. KSII Trans Internet Inf Syst, 2024, 18(7): 1888-906. doi：10.3837/tiis.2024.07.009

[1]	庞金龙, 赵新丽, 张振, 王豪杰, 周星琦, 杨玉梅, 李姗姗, 常小强, 李锋, 李娴. 皮肤黑色素瘤中MMRN2高表达促进肿瘤细胞的侵袭和迁移并与不良预后相关[J]. 南方医科大学学报, 2025, 45(7): 1479-1489.
[2]	赵倩, 张振, 周星琦, 荣翔宇, 刘雪柔, 赵新丽, 王豪杰, 庞金龙, 李姗姗, 李娴. 川芎增强替莫唑胺对小鼠脑转移黑色素瘤脑的抑制作用[J]. 南方医科大学学报, 2024, 44(6): 1088-1097.
[3]	石文惠, 刘小莲, 张贵明, 叶林萱, 周润华, 李亦蕾, 余乐. RITA体外选择性抑制BAP1缺失的皮肤黑色素瘤细胞的生长[J]. 南方医科大学学报, 2023, 43(5): 710-717.
[4]	李娟, 王爱莲, 李宁, 祝英泽, 李坤, 刘浩, 高自清. Aumolertinib可体内外抑制人脉络膜黑色素瘤MUM-2B细胞的增殖[J]. 南方医科大学学报, 2022, 42(11): 1604-1610.
[5]	钟友闻, 车文刚, 高盛祥. 轻型多尺度黑色素瘤目标检测网络模型的建立：基于注意力机制调控[J]. 南方医科大学学报, 2022, 42(11): 1662-1671.
[6]	刘忠强, 钟涛, 曹晓欢, 张煜. 基于组织修复的脑肿瘤图像配准方法[J]. 南方医科大学学报, 2021, 41(2): 292-298.
[7]	罗昕, 任重鲁, 刘小莲, 张贵明, 黄思思, 余乐, 李亦蕾. 靶向SF3B1突变型葡萄膜黑色素瘤的药物筛选[J]. 南方医科大学学报, 2021, 41(12): 1835-1842.
[8]	杨鑫, 李学妍, 张晓婷, 宋凡, 黄思娟, 夏云飞. 基于自适应Unet网络的鼻咽癌放疗危及器官自动分割方法[J]. 南方医科大学学报, 2020, 40(11): 1579-1586.
[9]	刘严友葓，徐虹铃，赖楠，杨子科，康世均 . 过表达IL-12的恶性黑色素瘤细胞在肿瘤免疫微环境重建过程中抑制T细胞表面PD-1的表达[J]. 南方医科大学学报, 2020, 40(06): 856-863.
[10]	邓旋，兰天俊，张明慧，陈之锋，陶谦，卢振泰. 基于局部灰度差异的快速自适应活动轮廓模型分割腮腺导管[J]. 南方医科大学学报, 2018, 38(12): 1485-.
[11]	黄奕晖，冯前进. 基于三维全卷积DenseNet的脑胶质瘤MRI分割[J]. 南方医科大学学报, 2018, 38(06): 661-.
[12]	门颖丽，康巧珍，丁聪，刘诗梦，江慧，王晓东，汲振余，刘鑫，王婷. 蛋白4.1家族在小鼠黑色素瘤细胞中的表达及其对细胞增殖的影响[J]. 南方医科大学学报, 2016, 36(05): 649-.
[13]	温锐，陈宏文，张雷，卢振泰. 基于引导滤波的多图谱医学图像分割[J]. 南方医科大学学报, 2015, 35(09): 1263-.
[14]	张雷，张明慧，卢振泰，冯前进，陈武凡. 基于多权重概率图谱的脑部图像分割[J]. 南方医科大学学报, 2015, 35(08): 1143-.
[15]	李欢，袁世梅，杨敏，段亮，王海燕，查何，李雪茹，孙晖，翁亚光，罗进勇，何通川，李崇雁，王嫣，李发琪，王志彪，周兰. 高强度聚焦超声抑制小鼠黑色素瘤细胞B16-F10在体内的转移[J]. 南方医科大学学报, 2015, 35(02): 223-.

SG-UNet：基于全局注意力和自校准卷积增强的黑色素瘤分割模型

SG-UNet: a melanoma segmentation model enhanced with global attention and self-calibrated convolution

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 14

参考文献 36

相关文章 15

编辑推荐

Metrics

本文评价