Diagnostic Value Of Texture Analysis For BI-RADS Grade 4 Or Above Calcification Based On Mammography Image

Objective:To investigate the feasibility of texture analysis based on breast X-ray image in distinguishing the benign and malignant lesions of suspected breast calcification.Materials and methods:Were retrospectively analyzed in October 2016 to March 2020 in the first affiliated hospital of Zheng Zhou university through regular breast calcifications found digital X-ray photography(BI-RADS classification into 4 a and above)after hospitalization,guided by X-ray godet positioning parallel resection for cases patients with pathological results of detailed data,all of the patients are women,and clinical did not hit a specific mass.There were 131 patients in the experimental group,including 136 calcification lesions,containing 98 benign lesions and 38 malignant lesions.There were 47 patients in the validation group,including 47 calcified lesions,containing 27 benign lesions and 20 malignant lesions.The benign group was 25-68 years old,with a mean age of about(45.1 ±0.817)years old and a median age of about 47.0 years old.The malignant group was 27-87 years old,with a mean age of about(47.6±1.712)years old and a median age of about 47.5 years old.Using a software named MaZda,extracted 300 texture features from these X-ray images of benign and malignant lesions,after by Fisher method(Fisher),minimization of both classification error probability and average correlation coefficients(POE+ACC),mutual information coefficients(MI)measure method to screen 10 best texture features,and incorporating three methods used to get the best combination of texture feature(MI+PA+F).Principal component analysis(PCA),linear discriminant analysis(LDA)and nonlinear discriminant analysis(NDA)were used to classify the four optimal texture features.The LDA and NDA were classified by K-NN(K-Nearest Neighbour)and ANN(artificial neutral network),respectively.The minimum error rate of 4 groups of texture features was analyzed for differentiating benign and malignant lesions.Then,the differences of 30 optimal texture features between benign and malignant lesions were compared by univariate and multivariate methods.The ROC curves for differentiating benign and malignant lesions were drawn,the AUC was calculated,and the diagnostic efficiency was evaluated and the independent risk factors affecting the disease were found.The statistical methods included chi-square test,two independent samples t test or Mann-Whitney U rank sum test,Logistic multivariate regression analysis and Receiver operating characteristic curve(ROC).Results:1.Texture features with statistically significant differences between the two groups of lesion images and their diagnostic efficiency:Lesions in the 2 groups,a total of 19 differences have statistical significance between the texture feature of the total variance of S(5,5),the total variance S(5,5),the total variance S(4,4),the total variance(3,3),the sum of squares S(4,4),entropy(1,1)and WavEnHL-s-4 better diagnostic performance(AUC=0.700,0.702,0.700,0.716,0.717,0.724,0.716),the texture feature is mostly are chosen from each part of gray level co-occurrence matrix method.There was no significant difference in AUC among the 6 subjects(all P=>0.05).2.Multi factor regression method was used to search for independent risk factors affecting the disease:Logistic binomial classification regression results indicated that the sum variance S(0,4),sum variance S(3,3),sum variance S(0,5),sum variance S(5,5)and difference entropy S(3,3)of texture parameters could be used as independent factors affecting the judgment of benign and malignant(P<0.05).3.Comparison of the error rate of the two groups of disease classification by texture analysis:for the three groups of the best texture features,the NDA/ANN-MI method had the lowest error rate,which was 16.18%(22/136);For MPF,the classification error rate of NDA/ANN-MPF was the lowest,which was 8.82%(12/136).There was no significant difference in the error rate between NDA/ANN-MI method and NDA/ANN-MPF method(χ2=3.361,P>0.05).4.In the validation group,the NDA/ANN-MPF classification method was used to judge the benign and malignant lesions of BI-RADS grade 4 or above,and the misdiagnosis rate was 4/47(8.51%).The accuracy,sensitivity,specificity,misdiagnosis rate,missed diagnosis rate,positive predictive value and negative predictive value of differential diagnosis were about 91.5%,85.0%,96.3%,3.7%,15.0%,94.4%and 89.7%respectively.Conclusions:The texture analysis based on mammography image can effectively distinguish the benign and malignant of BI-RADS grade 4 or above calcifications,which provides objective basis for the qualitative diagnosis of clinical calcifications.