The Value Of Machine Learning In The Differentiation Of Benign And Malignant Breast MRI Non-mass Enhancement Lesions

Part I Research on radiomics models based on threshold segmentation method in differentiating benign and malignant breast non-mass enhancement lesionsPurpose This study aims to compare the diagnostic performance of four radiomics models based on segmentation methods of signal intensity percentage thresholds（30%,40%,50%,60%）in differentiation of benign and malignant non-mass enhancement（NME）lesions on the first post-contrast sequence（S1）of breast MRI,and to obtain a better threshold,which lays the foundation for further research.Methods All patients with non-mass enhancement lesions in the Second Affiliated Hospital of Nantong University from January 2014 to September 2021 were retrospectively collected,and all enrolled cases were confirmed by surgery or biopsy pathology and at least 24 months of follow-up.Observer A used LIFEx to segment the area where the lesion was located on the second sequence of the dynamic contrast enhancement T1WI image to generate an initial three-dimensional volume of interest（VOI）covering the entire lesion,and then used the threshold calculation function of the software to retain the pixel signal intensity in the initial VOI>30%、40%、50%and 60%of the maximum signal intensity of the area,as 4 different final VOI(represented as VOI₃₀,VOI₄₀,VOI₅₀,VOI₆₀).Observer A segmented the lesions again after a 1-month interval,and two other observers B and C independently segmented the lesions,respectively,and extracted the radiomics features of all lesions after segmentation by 4 thresholds.Intra-and interclass correlation coefficients（ICCs）were used to evaluate the consistency of radiomics features extracted by physicians using threshold method to segment VOI,and the features with ICC>0.75 were retained.L1-based、F-test etc.methods were used for retaining important features and simplifying the models.Machine learning classifiers including Logistic Regression（LR）,Support Vector Machine（SVM）,Linear Support Vector Machine（LSVM）,Decision Tree（DT）and Random Forest（RF）were used to establish radiomics models for benign and malignant NME lesions classification,and a 10-time 5-fold cross-validation method was used to improve the stability of these models.The receiver operating characteristic（ROC）curve was used to evaluate the diagnostic performance of the models obtained by different threshold segmentation methods.The De Long test was used to compare the differences in the area under the ROC curve（AUC）of the models.Results A total of 154 patients with breast NME lesions（154lesions）were collected,including 79 benign lesions and 75 malignant lesions.They were all female patients,aged 24-83 years,with an average age of 46.3±14.3 years.4 different thresholds were used to segment NME lesions to obtain VOI₃₀,VOI₄₀,VOI₅₀ and VOI₆₀ and extract the radiomic features.Features with ICC≥0.75 accounted for>90%,and features with ICC≥0.90accounted for>75%.Regardless of the threshold,there was no significant difference in the proportion of features between the two segmentations of observer A and the sequential segmentation of three observers（P>0.05）.The VOI₃₀ and VOI₆₀ segmentation models obtained the highest AUC in the validation set（0.76 and 0.68,respectively）when selected the SVM classifier compared to other classifiers.The AUC was the highest（0.75）when the LSVM classifier was selected for the VOI₄₀segmentation model,and the AUC was the highest（0.78）when the LR classifier was selected for the VOI₅₀ segmentation model.The accuracy of the benign and malignant classification model established based on the 4threshold segmentation methods of VOI₃₀,VOI₄₀,VOI₅₀ and VOI₆₀ was0.70,0.66,0.71 and 0.65,respectively,the sensitivity was 0.66,0.71,0.68and 0.63,and the specificity was 0.74,0.62,0.73 and 0.67.The AUC between the groups was compared,only the AUC of the VOI₅₀ and VOI₆₀segmentation model had statistical differences（P=0.04）,and the AUCs of the other groups had no statistical differences（P>0.05）.Among the four models,the VOI₅₀ segmentation model which selected the LR classifier had the best diagnostic performance,which was significantly better than the VOI₆₀ segmentation model,and comparable to the VOI₃₀ and VOI₄₀segmentation models,but showed the highest accuracy,higher sensitivity and specificity.Conclusion Among the radiomics models established based on four different percentage thresholds to segment NME lesions for benign and malignant classification,the VOI₅₀ segmentation model which selects the LR classifier can better reflect the internal characteristics of the lesions,and the 50%threshold method can achieve fast and effective segmentation of breast NME lesions.Part II Research on multimodal features based on machine learning in differentiating benign and malignant breast nonmass enhancement lesionsPurpose This study aims to extract the radiomic features of NME lesions on the first（S1）and fifth post-contrast sequences（S5）,and combine patients’ clinical and routine MR characteristics（CRMC）to establish multimodal machine learning models.Then compare the diagnostic performance of the five models S1,S5,（S1+S5）,（S1+CRMC）,（S1+S5+CRMC）for benign and malignant NME lesions classification,so as to guide the decision-making of clinical diagnosis and treatment.Methods The patient data of the training and validation set were the same as those of the first part,and the 10-time 5-fold cross-validation method was also used,that is,the 154 samples were randomly divided into training set（123 cases）and validation set（31 cases）by 4:1,and the optimal result was selected after 10 rounds of repetition.In addition,35 patients with non-mass enhancement lesions in breast MR examination were retrospectively collected as the test set,and all enrolled cases were confirmed by surgery or biopsy pathology,including 19 patients from Nantong Third People’s Hospital from January 2020 to November 2021,and 16 patients from the Second Affiliated Hospital of Nantong University from October to December 2021.The clinical information of all patients was collected,including age,menopause status and clinical symptoms.Routine MR characteristics included type of breast fibroglandular tissue（FGT）,type of breast background parenchymal enhancement（BPE）,the largest diameter of NME lesions,distribution of NME lesions,internal enhancement features,type of time signal intensity curve,lowest mean ADC value,positive axillary lymph nodes,and presence of peritumoral edema.Referring to the optimal segmentation threshold（50%）obtained based on the S1 sequence in the first part,NME lesions were segmented on the S5 sequence images,and two radiomics models which selected the LR classifier based on the S5 sequence and（S1+S5）dual-sequence images were established.The diagnostic performance of radiomics models based on S1 sequence,S5 sequence and（S1+S5）dual-sequence was compared in the validation set;the patients’ CRMC was added to the radiomics model based on S1 sequence and（S1+S5）dual-sequence respectively to establish combination models.The diagnostic performance of the two new models was compared in the validation set,too.The diagnostic performance of all five models was evaluated in the test set.The comparison of measurement data between the two groups was performed by independent-samples t-test,and the comparison of categorical data was performed by chi-square test.The receiver operating characteristic（ROC）curve was used to evaluate the diagnostic performance of the post-enhanced single-sequence and dualsequence models and the combined model added with CRMC.De Long test was used to compare the area under the ROC curve（AUC）of models above.Results A total of 154 patients with breast NME lesions（154 lesions）in the training and validation set were collected,all female patients,aged 24-83 years,with an average age of 46.3±14.3 years.A total of 35 female patients in the test set were collected.They were 29-90 years old,with an average age of（51.4±15.8）years old.Among the clinical information and routine MR features of all patients（including training set,validation set and test set）,there were no significant differences in the type of breast fibroglandular tissue（FGT）,the largest diameter of NME lesions,the distribution of lesions,whether the axillary lymph nodes were positive or not,and whether there was peritumoral edema between the benign and malignant groups（P=0.721,0.248,0.461,0.203,0.069）.The mean age of patients in the malignant group was greater than that in the benign group（P<0.001）.The proportion of postmenopausal women in the malignant group was higher than that in the benign group（P<0.001）.There were significant differences in clinical symptoms,background parenchymal enhancement（BPE）type,internal enhancement characteristics and TIC type composition ratio between the two groups of benign and malignant patients（P ≤ 0.001）.Malignant lesions had higher-grade TIC types（P<0.001）,and lower mean ADC values than benign lesions（P<0.001）.According to univariate Logistic regression analysis,age,menopause status,clinical symptoms,BPE,internal enhancement characteristics,TIC type,and the lowest average ADC value were potential factors affecting the classification of benign and malignant NME lesions.Age,internal enhancement characteristics,TIC type,and lowest mean ADC value were significantly associated with breast cancer according to multivariate logistic regression analysis.Based on the 50% threshold method,the radiomics model of S1 sequence images,S5 sequence images and（S1+S5）combination sequences had no statistically significant difference in the classification performance of benign and malignant（AUC=0.78,0.71,0.80,P>0.05）.Five indicators of age,menopause status,internal enhancement characteristics,TIC type,and the lowest average ADC value were incorporated into the radiomics model whose AUC>0.75 based on S1 sequence and（S1+S5）dual-sequence,respectively.The AUC of the obtained（S1+CRMC）and（S1+S5+CRMC）model was 0.89 and 0.93 in the validation set respectively,between whom had no statistical difference（P=0.063）.However,the performance of the（S1+CRMC）model and（S1+S5+CRMC）model was significantly improved compared with the radiomics model alone（P<0.001）.The test set showed that the（S1+S5+CRMC）model had the best diagnostic performance（AUC=0.84）,which was better than the other four models of S1,S5,（S1+S5）,（S1+CRMC）（AUC=0.73,0.69,0.76,0.80）.Conclusion The radiomics features extracted from the postenhanced early and delayed phase images based on 50% threshold segmentation method have certain value in the differential diagnosis of NME lesions,the multimodal machine learning model based on postenhanced early and delayed radiomic features combined with clinical and routine MR characteristics can better distinguish benign and malignant breast NME lesions,and provide decision support for clinical diagnosis and treatment of breast cancer.