Classification of BI-RADS 3-5 Breast Lesions Based on MRI Radiomics

doi:10.11938/cjmr20222971

Abstract

Abstract:

The classification of the breast imaging-reporting and data system (BI-RADS) based on magnetic resonance imaging (MRI) refers to the classification of the degree of lesions according to the image signs of lesions, which is usually subjective. Moreover, the benign and malignant lesions of BI-RADS 3-5 are overlapping, which is prone to unnecessary invasive treatment due to high diagnostic categories in clinical diagnosis. To address these problems, this research applied radiomics for feature extraction and fusion of T₁-weighted (T1W) and dynamic contrast-enhanced (DEC) MRI. The least absolute shrinkage and selection operator (LASSO) algorithm was used to screen out the optimal feature collection of each type of MR image. Support vector machine (SVM), random forest (RF), K-nearest neighbour (KNN) and logistic regression (LR) algorithms were applied for BI-RADS 3-5 classification, based on which the benign and malignant lesions were further classified. The results showed that the classification accuracy of breast BI-RADS 3-5 by four radiomics models based on feature fusion was 81.25%, 87.50%, 78.38%, and 81.25%, respectively. Their accuracy in distinguishing the benign and malignant breast lesions was 90.91%, 93.55%, 92.73%, and 94.55%, respectively. This indicates that the combination of radiomics and machine learning correlation algorithm has a good effect on breast MRI BI-RADS classification and benign and malignant differentiation, and feature fusion can further improve the accuracy of classification prediction.

Key words: magnetic resonance imaging, breast lesions, radiomics, BI-RADS, feature fusion, classification

CLC Number:

O482.53

HAN Bing,XU Jing,WANG Yuanjun,WANG Zhongling. Classification of BI-RADS 3-5 Breast Lesions Based on MRI Radiomics[J]. Chinese Journal of Magnetic Resonance, 2023, 40(1): 52-67.

Figures/Tables 16

Fig. 1

Fig. 2

Table 1

Table 2

Table 3

Table 4

Fig. 3

Fig. 4

Table 5

Table 6

Table 7

Table 8

The optimal fusion features and corresponding coefficients screened by LASSO algorithm

影像组学特征	系数	特征详情
长轴长度（original_shape_MajorAxisLength）	0.050888	度量肿瘤最长轴的长度
联合能量（original_glcm_JointEnergy）	-0.016239	度量图像纹理中相邻灰度变换稳定程度
最大概率（original_glcm_MaximumProbability）	-0.071362	描述图像中出现次数最多的纹理特征
小波变换LLH方向-小区域强调（wavelet-LLH_glszm_SmallAreaEmphasis）	0.009840	度量小尺寸区域的分布情况
大依赖低灰度级强调（wavelet-LLH_gldm_LargeDependenceLowGrayLevelEmphasis）	-0.014709	描述体素强相关关系与低阶灰度的联合分布情况
低灰度区域强调（wavelet-LHH_glszm_LowGrayLevelZoneEmphasis）	-0.002699	描述低灰度级区域体素的分布
区域熵（wavelet-LHH_glszm_ZoneEntropy）	0.090181	度量灰度区域大小与灰度级分布的不稳定性
相关性信息测度2（wavelet-HHL_glcm_Imc2）	-0.086920	量化纹理的复杂性
小波变换HHL方向-小区域强调（wavelet-HHL_glszm_SmallAreaEmphasis）	-0.051711	度量小区域尺寸的分布情况
小依赖低灰度级强调（wavelet-HHL_gldm_SmallDependenceLowGrayLevelEmphasis）	-0.012360	描述体素小相关关系与低阶灰度的联合分布情况
中位数（wavelet-HHH_firstorder_Median）	-0.023237	描述肿瘤区域的灰度中位数
逆方差（wavelet-HHH_glcm_InverseVariance）	0.013141	度量肿瘤图像局部均匀程度
粗糙度（wavelet-LLL_ngtdm_Coarseness）	-0.014267	度量中心体素与其邻域体素灰度值的平均差异
短游程低灰度级强调（square_glrlm_ShortRunLowGrayLevelEmphasis）	-0.048257	度量低阶灰度值与短游程长度的联合分布情况
强度（square_ngtdm_Strength）	-0.002079	度量肿瘤图像的灰度变化程度
第10%分位值（squareroot_firstorder_10Percentile）	-0.003926	指肿瘤区域10%分位数的灰度值
小区域低灰度级强调（exponential_glszm_SmallAreaLowGrayLevelEmphasis）	0.003075	描述低灰度小尺寸区域体素的分布
小依赖强调（logarithm_gldm_SmallDependenceEmphasis）	-0.002973	描述与领域内体素相关性较小的体素分布情况

Table 8

Fig. 5

Fig. 6

Table 9

Table 10

References 28

[1]	SIEGEL R L, MILLER K D, FUCHS H E, et al. Cancer statistics, 2021[J]. CA: A Cancer J Clin, 2021, 71(1): 7-33. doi: 10.3322/caac.21654
[2]	FENG R M, ZONG Y N, CAO S M, et al. Current cancer situation in China: good or bad news from the 2018 global cancer statistics?[J]. Cancer Commun, 2019, 39(1): 22-33. doi: 10.1186/s40880-019-0368-6
[3]	VALDORA F, HOUSSAMI N, ROSSI F, et al. Rapid review: radiomics and breast cancer[J]. Breast Cancer Res Treat, 2018, 169(2): 217-229. doi: 10.1007/s10549-018-4675-4
[4]	MANN R M, CHO N, MOY L. Breast MRI: State of the art[J]. Radiology, 2019, 292(3): 520-536. doi: 10.1148/radiol.2019182947 pmid: 31361209
[5]	WAKANA M, HYUNG W C, JOINES M M, et al. Quantitative predictors of response to neoadjuvant chemotherapy on dynamic contrast-enhanced 3T breast MRI[J]. J Breast Imaging, 2022, 4(2): 168-176. doi: 10.1093/jbi/wbab095
[6]	SHAHAN C L, LAYNE G P. Advances in breast imaging with current screening recommendations and controversies[J]. Obstet Gynecol Clin North Am, 2022, 49(1): 1-33. doi: 10.1016/j.ogc.2021.11.001
[7]	LIU Y, CHEN J C, HU X Y, et al. Classification and localization of meningioma and acoustic neuroma in cerebellopontine angle based on mask RCNN[J]. Chinese J Magn Reson, 2021, 38(1): 58-68.
	刘颖, 陈静聪, 胡小洋, 等. 基于Mask RCNN的桥小脑角区脑膜瘤与听神经瘤分类定位研究[J]. 波谱学杂志, 2021, 38(1): 58-68.
[8]	WEAVER O, LEUNG J W T. Biomarkers and imaging of breast cancer[J]. AJR Am J Roentgenol, 2018, 210(2): 271-278. doi: 10.2214/AJR.17.18708
[9]	RAWASHDEH M, LEWIS S, ZAITOUN M, et al. Breast lesion shape and margin evaluation: BI-RADS based metrics understate radiologists' actual levels of agreement[J]. Comput Biol Med, 2018, 96(1): 294-298. doi: 10.1016/j.compbiomed.2018.04.005
[10]	SATAKE H, ISHIGAKI S, ITO R, NAGANAWA S. Radiomics in breast MRI: current progress toward clinical application in the era of artificial intelligence[J]. Radiol Med. 2022, 127(1): 39-56. doi: 10.1007/s11547-021-01423-y
[11]	LEITHNER D, HORVAT J V, OCHOA-ALBIZTEGUI R E, et al. Imaging and the completion of the omics paradigm in breast cancer[J]. Der Radiologe, 2018, 58(1): 7-13. doi: 10.1007/s00117-018-0409-1
[12]	WANG N, WANG Y J, LIAN P. Prediction of preoperative T staging of rectal cancer based on radiomics[J]. Chinese J Magn Reson, 2022, 39(1): 43-55.
	王楠, 王远军, 廉朋. 基于影像组学的直肠癌术前T分期预测[J]. 波谱学杂志, 2022, 39(1): 43-55.
[13]	IMBRIACO M, CUOCOLO R. Does texture analysis of MR images of breast tumors help predict response to treatment?[J]. Radiology, 2018, 286(2): 421-423. doi: 10.1148/radiol.2017172454 pmid: 29356631
[14]	YIN X X, HADJILOUCAS S, ZHANG Y C, et al. MRI radiogenomics for intelligent diagnosis of breast tumors and accurate prediction of neoadjuvant chemotherapy responses — A review[J]. Comput Meth Prog Bio, 2021, 214: 106510. doi: 10.1016/j.cmpb.2021.106510
[15]	NARANJO I D, GIBBS P, REINER J S, et al. Radiomics and machine learning with multiparametric breast MRI for improved diagnostic accuracy in breast cancer diagnosis[J]. Diagnostics, 2021, 11(6): 919-932. doi: 10.3390/diagnostics11060919
[16]	PRATIKSHA Y, SURBHI C. Effectivity of combined diffusion-weighted imaging and contrast-enhanced MRI in malignant and benign breast lesions[J]. Pol J Radiol, 2018, 83: e82-e93.
[17]	HU Q, WHITNEY H M, GIGER M L. Radiomics methodology for breast cancer diagnosis using multiparametric magnetic resonance imaging[J]. J Med Imaging, 2020, 7(4): 044502.
[18]	YE D M, WANG H T, YU T. The application of radiomics in breast MRI: A review[J]. Technol Cancer Res Tr, 2020, 19: 1-16.
[19]	PAREKH V S, JACOBS M A. Multiparametric radiomics methods for breast cancer tissue characterization using radiological imaging[J]. Breast Cancer Res Tr, 2020, 180(2): 407-421. doi: 10.1007/s10549-020-05533-5 pmid: 32020435
[20]	DEMIRCIOGLU A, GRUENEISEN J, INGENWERTH M, et al. A rapid volume of interest-based approach of radiomics analysis of breast MRI for tumor decoding and phenotyping of breast cancer[J]. PLoS ONE, 2020, 15(6): e0234871.
[21]	TSAROUCHI M I, VLACHOPOULOS G F, KARAHALIOU A N, et al. Multi-parametric MRI lesion heterogeneity biomarkers for breast cancer diagnosis[J]. Phys Medica, 2020, 80(2): 101-110. doi: 10.1016/j.ejmp.2020.10.007
[22]	ZHANG Q, PENG Y S, LIU W, et al. Radiomics based on multimodal MRI for the differential diagnosis of benign and malignant breast lesions[J]. J Magn Reson Imaging, 2020, 52(2): 596-607. doi: 10.1002/jmri.27098 pmid: 32061014
[23]	HAO W, GONG J, WANG S P, et al. Application of MRI radiomics-based machine learning model to improve contralateral BI-RADS 4 lesion assessment[J]. Front Oncol, 2020. doi: 10.3389/fonc.2020.531476. doi: 10.3389/fonc.2020.531476
[24]	QIAO M Y, LI C K, SUO S T, et al. Breast DCE-MRI radiomics: a robust computer-aided system based on reproducible BI-RADS features across the influence of datasets bias and segmentation methods[J]. Int J Comput Ass Rad, 2020, 15(5): 921-930.
[25]	SCAPICCHIO C, GABELLONI M, BARUCCI A, et al. A deep look into radiomics[J]. La Radiologia Medica, 2021, 126: 1296-1311. doi: 10.1007/s11547-021-01389-x
[26]	DOMINGUES I, ABREU P H, SANTOS J. Bi-Rads classification of breast cancer: A new pre-processing pipeline for deep models training[C]// IEEE International Conference of Image Processing (ICIP). IEEE, 2018.1378-1382.
[27]	SIDDEEQ S, LI J, BHATTI H M A, et al. Deep learning RN-BCNN model for breast cancer BI-RADS classification[C]. ICIGP 2021: 2021 The 4th International Conference on Image and Graphics Processing. 2021: 219-225.
[28]	WU P Q. Research progress of radiomics in lymph node metastasis of breast cancer[J]. Journal of Molecular Imaging, 2020, 43 (1): 31-35.
	吴佩琪. 影像组学在乳腺癌淋巴结转移中的研究进展[J]. 分子影像学杂志, 2020, 43(1): 31-35.

特征类别	特征描述	特征数	特征输入
shape	表面积、伸长率、平面度、球形度、最大二维直径等	14	原始图像
firstorder	能量、熵、峰值、最大值、平均值、方差、第10百分位数、第90百分位数等	342	原始图像、派生图像
GLCM	自相关、联合平均、聚类突出度、聚类阴影、对比、相关、差熵等	418	原始图像、派生图像
GLRLM	长游程强调、短游程强调、灰度不均匀性、游程百分比、游程方差、游程熵等	304	原始图像、派生图像
GLDM	小依赖度、大依赖度、依赖不均匀性、依赖方差、依赖熵等	266	原始图像、派生图像
GLSZM	小区域强调、大区域强调、灰度不均匀性、区域百分比、区域方差、区域熵等	304	原始图像、派生图像
NGTDM	粗糙度、对比度、繁忙度、复杂度、强度	95	原始图像、派生图像

影像组学特征	系数	特征详情
联合熵（original_glcm_JointEntropy）	0.010878	度量邻域强度值的可变性
平均值（log-sigma-5-0-mm-3D_firstorder_Mean）	0.120571	描述肿瘤区域的平均灰度值
小区域低灰度级强调（wavelet-HLH_glszm_SmallAreaLowGrayLevelEmphasis）	-0.110678	描述低灰度小尺寸区域体素的分布
归一化相关不均匀性（wavelet-HHL_gldm_DependenceNonUniformityNormalized）	-0.089366	描述GLDM中体素相关关系的相似程度
依赖方差（wavelet-HHL_gldm_DependenceVariance）	0.007990	描述GLDM中依赖大小的方差
相关性信息测度2（wavelet-HHH_glcm_Imc2）	-0.045300	量化纹理的复杂性
游程方差（wavelet-HHH_glrlm_RunVariance）	-0.059722	度量游程长度的方差
依赖熵（gradient_gldm_DependenceEntropy）	0.019160	度量GLDM中依赖大小与灰度级分布的随机性程度
小依赖性低灰度级强调（gradient_gldm_SmallDependenceLowGrayLevelEmphasis）	-0.021736	描述体素小相关性与低阶灰度值的联合分布情况
相关性信息测度1（squareroot_glcm_Imc1）	0.030811	量化纹理的复杂性
强度（exponential_ngtdm_Strength）	-0.055385	度量肿瘤图像的灰度变化程度

影像组学特征	系数	特征详情
原始图像最小值（original_firstorder_Minimum）	-0.013973	描述肿瘤区域的最小灰度值
大依赖高灰度级强调（original_gldm_LargeDependenceHighGrayLevelEmphasis）	0.039997	测量具有较高灰度值和体素强相关关系的联合分布情况
偏度（wavelet-LHH_firstorder_Skewness）	0.001540	度量强度平均值的分布不对称性
相关性信息测度1（wavelet-LHH_glcm_Imc1）	0.057154	量化纹理的复杂性
小波变换LLL方向联合熵（wavelet-LLL_glcm_JointEntropy）	0.197864	度量邻域强度值的可变性
归一化逆差矩（gradient_glcm_Idmn）	0.034171	度量肿瘤图像的局部平均程度
平方变换最小值（square_firstorder_Minimum）	0.037978	描述肿瘤区域的最小灰度值
平方变换联合熵（square_glcm_JointEntropy）	0.064132	度量邻域强度值的可变性
大区域低灰度级强调（squareroot_glszm_LargeAreaLowGrayLevelEmphasis）	-0.014832	测量图像中具有较低灰度值的较大尺寸区域体素的分布
低灰度区域强调（logarithm_glszm_LowGrayLevelZoneEmphasis）	-0.055083	描述低灰度级区域体素的分布
低灰度强调（logarithm_gldm_LowGrayLevelEmphasis）	-0.084834	描述低阶灰度体素的分布情况

影像组学特征	系数	特征详情
最大2D直径（列）（original_shape_Maximum2DDiameterColumn）	0.081848	冠状平面中肿瘤表面网格顶点之间最大的欧几里得距离
相关性信息测度1（wavelet-LHL_glcm_Imc1）	0.018552	量化纹理的复杂性
集群阴影（wavelet-LHH_glcm_ClusterShade）	-0.022909	描述图像的褶皱程度
小波变换HLL方向粗糙度（wavelet-HLL_ngtdm_Coarseness）	-0.051320	度量中心体素与其邻域体素之间平均差异
小区域低灰度级强调（wavelet-HLH_glszm_SmallAreaLowGrayLevelEmphasis）	-0.031217	描述图像中具有较低灰度值的较小尺寸区域体素分布
相关性信息测度2（wavelet-HHL_glcm_Imc2）	-0.066184	量化纹理的复杂性
逆方差（wavelet-HHH_glcm_InverseVariance）	0.010646	度量肿瘤图像局部均匀程度
归一化灰度不均匀度（wavelet-LLL_glszm_GrayLevelNonUniformityNormalized）	-0.138090	测量图像中灰度强度值的可变性
小波变换LLL方向粗糙度（wavelet-LLL_ngtdm_Coarseness）	-0.086469	度量中心体素与其邻域体素之间平均差异
归一化区域大小不均匀度（square_glszm_SizeZoneNonUniformityNormalized）	-0.022937	测量整个图像中大小区域体积的可变性
最小值（exponential_firstorder_Minimum）	-0.038634	描述肿瘤区域的最小灰度值
对比度（logarithm_ngtdm_Contrast）	-0.019143	度量体素灰度的空间变化率

特征	模型	准确率	Kappa系数	海明损失	Micro AUC	Macro AUC
平扫特征	SVM	64.86%	0.419	0.297	0.848	0.820
	RF	70.27%	0.421	0.333	0.878	0.846
	KNN	70.27%	0.421	0.333	0.830	0.769
	LR	71.43%	0.421	0.333	0.858	0.882
增强特征	SVM	70.27%	0.428	0.297	0.895	0.847
	RF	86.49%	0.428	0.297	0.928	0.862
	KNN	76.19%	0.445	0.297	0.875	0.828
	LR	78.57%	0.428	0.297	0.948	0.929
融合特征	SVM	81.25%	0.606	0.224	0.881	0.855
	RF	87.50%	0.676	0.188	0.951	0.988
	KNN	78.38%	0.606	0.224	0.886	0.867
	LR	81.25%	0.606	0.224	0.961	0.989