·28· 智能系统学报 第16卷 with applications,2012,64(5):1153-1162 [4]GINSBURG O,BRAY F,COLEMAN M P,et al.The global burden of women's cancers:a grand challenge in global health[J].The lancet,2017,389(10071):847-860. [5]MENEZES G L,KNUTTEL F M,STEHOUWER B L,et al.Magnetic resonance imaging in breast cancer:a literat- (c)YOLOV3 (d)YOLOV3(1) ure review and future perspectives[J].World journal of clinical oncology,2014,5(2):61-70. [6]袁红梅，余建群，褚志刚，等.动态增强MRI、超声及 X射线对乳腺良恶性病灶诊断的对比研究).中国普外 基础与临床杂志，2015,22(2)：246-250. YUAN Hongmei,YU Jianqun,CHU Zhigang,et al.Com- (e)YOLOV3(2) (f)YOLOV3 (3) parative study of dynamic contrast-enhanced breast MRI. 图12良性测试样本效果展示 ultrasound,and X-ray mammography in differential dia- Fig.12 Benign test sample effect display gnosis of benign and malignant breast lesions[J].Chinese 5结束语 journal of bases and clinics in general surgery,2015,22(2): 246-250. 针对传统乳腺超声肿瘤识别方法均采用人工 [7]中国抗癌协会乳腺癌专业委员会.中国抗癌协会乳腺癌 提取的特征逐步实现ROI区域定位和肿瘤分类 诊治指南与规范(2017年版)[J].中国癌症杂志， 往往识别精度低、鲁棒性较差且通用性不强，目 2017,27(9):695-759 前基于深度学习的方法又仅限于肿瘤ROI区域 Breast cancer professional committee of Chinese anti-can- 的定位或对给定的肿瘤ROI区域进行分类，本文 cer association.Guidelines and specifications for breast cancer diagnosis and treatment of China anti cancer associ- 提出采用深度学习中的YOLOv.3算法同时实现了 ation (2017 Edition)[J].China oncology,2017,27(9): 对良恶性的分类和肿瘤ROI区域的定位，同时针 695-759. 对乳腺肿瘤识别中的问题对算法进行了改进，经 [8]周星彤，沈松杰，孙强.中国乳腺癌筛查现状及进展[， 过实验证明，在引入Res2Net、DownSample模块 中国医学前沿杂志，2020,12(3)：6-11. 和残差密集网络后YOLOv3算法有着更高的检测 ZHOU Xingtong,SHEN Songjie,SUN Qiang.Current 精度，其在测试集上mAP达到0.7959，平均IOU situation and progress of breast cancer screening in 达到0.8259，相比于传统的YOLOv3算法分别提 China[J].Chinese journal of the frontiers of medical sci- 高了4.56%和2.35%，今后可进一步优化算法提 ence (electronic version),2020,12(3):6-11. 高检测精度。 [9]Cai L,Wang X,Wang Y,et al.Robust phase-based tex- 经专业医生鉴定，采用改进后的YOLOv3算 ture descriptor for classification of breast ultrasound im- 法不仅同时实现对肿瘤ROI区域的定位和良恶 ages[J].BioMedical Engineering OnLine,2015,14(1): 性的分类，而且取得了较好的检测效果，使得人 1-21 [10]HUANG Y L.JIANG Y R.CHEN D R,et al.Computer- 工智能应用更接近实际操作环境，有效提升基层 aided diagnosis with morphological features for breast le- 医生诊断能力，降低专科医生工作强度，有着极 sion on sonograms[J].Ultrasound in obstetrics and 大的应用价值。 gynecology,.2008,32(4):565-572. 参考文献： [11]KABIR S M,BHUIYAN M I H.Classification of breast tumour in contourlet transform domain[C]//2018 10th In- [1]CHEN Wanqing,ZHENG Rongshou,BAADE P D,et al. ternational Conference on Electrical and Computer Engin- Cancer statistics in China,2015[J].CA:a cancer journal eering (ICECE).Dhaka,Bangladesh,2018:289-292. for clinicians.2016,66(2):115-132. [12]MENON R V,RAHA P,KOTHARI S,et al.Automated [2]SIEGEL RL,MILLER K D,JEMAL A.Cancer statistics, detection and classification of mass from breast ultra- 2016[J].CA:a cancer journal for clinicians,2016,66(1): sound images[C]//2015 5th National Conference on Com- 7-30 puter Vision,Pattern Recognition,Image Processing and [3]LO C S,WANG C M.Support vector machine for breast Graphics.Patna,India,2015:1-4. MR image classification[J].Computers and mathematics [13]肖婷.基于深度特征迁移与融合的乳腺超声图像分5 结束语 针对传统乳腺超声肿瘤识别方法均采用人工 提取的特征逐步实现 ROI 区域定位和肿瘤分类 往往识别精度低、鲁棒性较差且通用性不强，目 前基于深度学习的方法又仅限于肿瘤 ROI 区域 的定位或对给定的肿瘤 ROI 区域进行分类，本文 提出采用深度学习中的 YOLOv3 算法同时实现了 对良恶性的分类和肿瘤 ROI 区域的定位，同时针 对乳腺肿瘤识别中的问题对算法进行了改进，经 过实验证明，在引入 Res2Net、DownSample 模块 和残差密集网络后 YOLOv3 算法有着更高的检测 精度，其在测试集上 mAP 达到 0.795 9，平均 IOU 达到 0.825 9，相比于传统的 YOLOv3 算法分别提 高了 4.56% 和 2.35%，今后可进一步优化算法提 高检测精度。 经专业医生鉴定，采用改进后的 YOLOv3 算 法不仅同时实现对肿瘤 ROI 区域的定位和良恶 性的分类，而且取得了较好的检测效果，使得人 工智能应用更接近实际操作环境，有效提升基层 医生诊断能力，降低专科医生工作强度，有着极 大的应用价值。 参考文献： CHEN Wanqing, ZHENG Rongshou, BAADE P D, et al. Cancer statistics in China, 2015[J]. CA: a cancer journal for clinicians, 2016, 66(2): 115–132. [1] SIEGEL R L, MILLER K D, JEMAL A. Cancer statistics, 2016[J]. CA: a cancer journal for clinicians, 2016, 66(1): 7–30. [2] LO C S, WANG C M. Support vector machine for breast MR image classification[J]. Computers and mathematics [3] with applications, 2012, 64(5): 1153–1162. GINSBURG O, BRAY F, COLEMAN M P, et al. The global burden of women’s cancers: a grand challenge in global health[J]. The lancet, 2017, 389(10071): 847–860. [4] MENEZES G L, KNUTTEL F M, STEHOUWER B L, et al. Magnetic resonance imaging in breast cancer: a literat￾ure review and future perspectives[J]. World journal of clinical oncology, 2014, 5(2): 61–70. [5] 袁红梅, 余建群, 褚志刚, 等. 动态增强 MRI、超声及 X 射线对乳腺良恶性病灶诊断的对比研究 [J]. 中国普外 基础与临床杂志, 2015, 22(2): 246–250. YUAN Hongmei, YU Jianqun, CHU Zhigang, et al. Com￾parative study of dynamic contrast-enhanced breast MRI, ultrasound, and X-ray mammography in differential dia￾gnosis of benign and malignant breast lesions[J]. Chinese journal of bases and clinics in general surgery, 2015, 22(2): 246–250. [6] 中国抗癌协会乳腺癌专业委员会. 中国抗癌协会乳腺癌 诊治指南与规范 (201 7 年版)[J]. 中国癌症杂志, 2017,27(9): 695–759. Breast cancer professional committee of Chinese anti-can￾cer association. Guidelines and specifications for breast cancer diagnosis and treatment of China anti cancer associ￾ation (2017 Edition)[J]. China oncology, 2017,27(9): 695–759. [7] 周星彤, 沈松杰, 孙强. 中国乳腺癌筛查现状及进展 [J]. 中国医学前沿杂志, 2020, 12(3): 6–11. ZHOU Xingtong, SHEN Songjie, SUN Qiang. Current situation and progress of breast cancer screening in China[J]. Chinese journal of the frontiers of medical sci￾ence (electronic version), 2020, 12(3): 6–11. [8] Cai L, Wang X, Wang Y, et al. Robust phase-based tex￾ture descriptor for classification of breast ultrasound im￾ages[J]. BioMedical Engineering OnLine, 2015, 14(1): 1–21. [9] HUANG Y L, JIANG Y R, CHEN D R, et al. Computer￾aided diagnosis with morphological features for breast le￾sion on sonograms[J]. Ultrasound in obstetrics and gynecology, 2008, 32(4): 565–572. [10] KABIR S M, BHUIYAN M I H. Classification of breast tumour in contourlet transform domain[C]//2018 10th In￾ternational Conference on Electrical and Computer Engin￾eering (ICECE). Dhaka, Bangladesh, 2018: 289–292. [11] MENON R V, RAHA P, KOTHARI S, et al. Automated detection and classification of mass from breast ultra￾sound images[C]//2015 5th National Conference on Com￾puter Vision, Pattern Recognition, Image Processing and Graphics. Patna, India, 2015: 1–4. [12] [13] 肖婷. 基于深度特征迁移与融合的乳腺超声图像分 (c) YOLOV3 benign 97.28% (d) YOLOV3 (1) benign 98.48% (e) YOLOV3 (2) benign 98.89% (f) YOLOV3 (3) benign 99.97% 图 12 良性测试样本效果展示 Fig. 12 Benign test sample effect display ·28· 智 能 系 统 学 报 第 16 卷