Preliminary Study On The Value Of The Radiomics Model Based On Multimodal MRI In Differentiating Benign From Malignant Breast Lesions

Part one.Study on the value of the radiomics model based on each sequence in differentiating benign from malignant breast lesionsObjective:To construct five radiomics models based on T2WI,T1WI,ADC,DKI and DCE-MRI respectively,and to evaluate the discriminatory ability of each radiomics model for benign and malignant breast lesions.Materials and Methods:Patients who underwent a preoperativebreast 3.0 T MRI examination(conventional MRI,DWI and DCE-MRI sequences)and then underwent biopsy or surgery at the second affiliated hospital of Soochow university were enrolled in this retrospective study between June 2015 and August 2019.According to the final pathological diagnosis results,the study cohort included 112 breast malignant lesions and 95 breast benign lesions.All the included cases were treated with the same MRI equipment,and the scanning sequence parameters were consistent.The DCE-MRI images of each case were imported into Omni-Kinetic software to obtain pharmacokinetic parameter maps.The DW images of each case were imported into FireVoxel software to obtain kurtosis maps and diffusivity maps.After each case was preprocessed,the eighth postcontrast images were used to sketch the three-dimensional ROI of the lesion.The location and size of the lesion was determined by DWI and T2WI images as reference and manually sketched layer by layer along the contour of the lesion,and then multilayer fusion was used to obtain the three-dimensional ROI of the lesion.The ROIs based on the eighth postcontrast images were mapped to the pharmacokinetic parameter maps,T1WI,T2WI,ADC and DKI images.First,the reproducibility of the observer segmentation was evaluated by evaluating the reproducibility of the feature extraction from the five sequences in this study;the ICC is greater than 0.75 for good reproducibility.Then,all cases were divided into the training set and test set in a ratio of 7:3 according to the examination time.Each lesion ROI voxel of training set were extracted the computer derivative features by Python 3.6 and Pyradiomics toolkit.There were three categories of features extracted,including shape features,first-order features and texture features.Using Spearman correlation test and LASSO algorithm selected features to determine the key feature set.The most valuable features were analyzed by logistic regression analysis.The support vector machine(SVM)classifier was used to build the radiomics model.SVM models with the key features were trained based on the training set,and then ROC curves were calculated in the individual validation sets to evaluate the model.The discriminative abilities of the five radiomics models were evaluated using ROC curve for identifying benign and malignant breast lesions,and the accuracies,sensitivities and specificities were calculated.Results:The study cohort included 207 breast lesions,including 95 benign and 112 malignant lesions.The most common malignant lesion was invasive ductal carcinoma,which was found in 86(76.79%)patients.Among the benign lesions,the most common lesion was fibroadenoma,which was found in 48(50.53%)patients.The ICC range of the twice images feature extraction of the observer A was 0.804-0.996.The ICC range of the observer A(the first)and the observer B was 0.751-0.996.The reproducibility between the inter-observer and intra-observer is good.A total of 862 radiomics features were obtained through the above multiple feature extraction methods,and finally the dimensions were reduced to 7,29,5,3 and 17 non-zero coefficient optimal diagnostic features of five sequences(T2WI,T1WI,ADC,DKI,pharmacokinetic parameter maps)respectively through the multi-step feature selection methods.Each radiomics model based on the training set has a reasonable diagnostic efficiency in the differentiation of breast benign and malignant lesions.In the test group,the AUC of each model(T2WI,T1WI,ADC,DKI,pharmacokinetic parameter maps)was 0.791、0.730、0.770、0.788 and 0.836,respectively;and the accuracy of each model was 0.708、0.646、0.813、0.729 and 0.750,respectively.Conclusion:This study show that there was a good reproducibility between intra-observer and inter-observer about feature extraction based on T2WI,T1WI,ADC,DKI and pharmacokinetic parameter maps.The radiomics feature models of the five individual sequences all showed the reasonable diagnostic efficiency,and the model based on pharmacokinetic parameter maps obtained the strongest ability to discriminate between benign and malignant breast lesions.Radiomics can be used as an auxiliary method to distinguish benign and malignant breast lesions before surgery.Part two.Study on the value of the radiomics model based on various combinations of five sequences in differentiating benign from malignant breast lesionsObjective:The second part of the study will establish the optimal radiomics model to distinguish breast benign and malignant lesions.The radiomics feature models of various combinations of the five sequences in the first part were built.The AUC of radiomics feature model of each combination of the five sequences was used to evaluate the diagnostic efficacy.The ability of the optimal radiomics model in the differential diagnosis of breast benign and malignant lesions was evaluated using the AUC,accuracy,sensitivity and specificity.Materials and Methods:207 female patients with 207 histopathology-confirmed breast lesions including 95 benign and 112 malignant were enrolled in our study.The first part of the study have obtained the radiomics models for the differential diagnosis of benign and malignant breast lesions based on T2WI,T1WI,ADC,DKI,pharmacokinetic parameter maps,respectively.In this part of the study,the radiomics feature models of various combinations of the five sequences in the first part were built.First,all cases were divided into the training set and test set in a ratio of 7:3 according to the examination time.For various combinations of five sequences,the selected features of the respective sequences were combined,reduced again with Spearman’s rank correlation coefficients and once more fed into the LASSO logistic regression model to choose the key features in the fusion feature set.SVM classifier was used to build the radiomics model based on the various combination of five sequence.SVM models with the key features were trained based on the training set,and then ROC curves were calculated in the individual validation sets to evaluate the model.Since both ADC map and DKI are derived from DWI sequences,this combination is not included in the discussion of two sequence combinations.For the models that combined features from two sequences(except ADC+DKI),the combination of features from any two of the three sequences,that is,T2WI,pharmacokinetic parameter maps and DKI,could improve the diagnostic efficiency of the radiomics model,but adding either features of T1W images or features of ADC maps to other the features of other sequences could not improve the diagnostic performance,so the features from T1W images and ADC maps were not included in the subsequent model hat combined feature sets from three sequences.When T2WI,DKI and pharmacokinetic parameter maps were combined,the radiomics signature model exhibited the highest AUC and obtained the best diagnostic accuracy.The above three models were all constructed in the training set and tested in the corresponding test set.The discriminative abilities of the radiomics models of various combination of five sequence were evaluated using ROC curve for identifying benign and malignant breast lesions,and the accuracies,sensitivities and specificities were calculated.The optimal radiomics signature model composed of twenty-two features.The Kolmogorov-Smirnov test(K-S test)was applied to examine the normality of the feature distribution of each optimal model.After the K-S test,all features in the optimal model,which were quantitative variables,followed a normal distribution,so comparisons between two groups(benign and malignant)were made with Student’s t-test.The distribution of the selected features was also described by the mean,median and variance values.All statistical tests were two-sided,and P values less than 0.05 were considered statistically significant.Results:This study built the radiomics feature models of various combinations of the five sequences in the first part and found the optimal radiomics model to distinguish breast benign and malignant lesions.In the test set,the AUCs of radiomics feature model based on combinations of two sequence(T2WI+T1WI,T2WI+ADC,T2WI+DKI,T2WI+pharmacokinetic parameter maps,T1WI+ADC,T1WI+DKI,T1WI+pharmacokinetic parameter maps,ADC+pharmacokinetic parameter maps,DKI+pharmacokinetic parameter maps)were 0.782,0.664,0.805,0.873,0.664,0.788,0.832,0.716 and 0.884,respectively.The AUC,accuracy,sensitivity and specificity of the optimal radiomics model(T2WI+DKI+pharmacokinetic parameter maps)were 0.921,0.833,0.750 and 0.950,respectively.22 features were selected in optimal radiomics signature model,including 12 texture features,9 first-order features,and 1 shape feature.In the optimal radiomics signature model,sixteen of twenty-two features were significantly different between benign and malignant breast lesions(p<0.05),including 10 texture features,6 first-order features.Eleven of sixteen features were significantly different between benign and malignant breast lesions(p<0.05)extracted from the pharmacokinetic parameter maps.Conclusions:This study shows radiomics features extracted from multiple MR sequences were useful in establishing a predictive model that could help differentiate between benign and malignant breast lesions.The radiomics features of quantitative DCE pharmacokinetic parameter maps played an important role in the differential diagnosis of breast lesions.T2W images and DKI were useful adjuncts to DCE-MRI for improving the diagnostic efficiency.The optimal radiomics feature model was based on T2WI,DKI and quantitative DCE pharmacokinetic parameter maps.Therefore,the radiomics signature model is helpful for the preoperative diagnosis and it is expected to become an objective and comprehensive new diagnostic method with the deepening of research.