The Study Of Radiomics And Deep Learning Based Contrast-Enhanced Spectral Mammography For Prediction Of Triple-Negative Breast Cancer And Pathological Features

Part ⅠValue of contrast-enhanced spectral mammography based radiomics and deep learning for prediction of triple-negative breast cancerBackgroundBreast cancer is one of the most frequently malignant cancer.Triple-negative breast cancer（TNBC）is a special molecular subtype of breast cancer with high invasion,aggressive proliferation and recurrence,worse prognosis.So,it’s important to diagnose TNBC timely and precisely.At present immunohistochemistry based on biopsy or surgery is the used for the diagnosis of TNBC.Biopsy is invasive.Furthermore,only part of tissue is acquired with the method of biopsy.Given the heterogeneity of tumor,the reliability of diagnostic result is limited.A noninvasive and reliable mothed for the diagnosis of TNBC is urgently needed.Radiomics and deep learning was widely used in the diagnosis of malignant tumor and its subtype.The novel contrast-enhanced spectral mammography（CESM）includes the low-energy images which show the morphological characteristics and the subtracted images which highlight neovascularization characteristics in breast disease.CESM was widely used in radiomics and deep learning analysis.However,only few previous studies focused on the prediction of TNBC based on CESM which with small patient datasets,lack of testing set,and too little metrics in the valuation of prediction model.The results in these studies were not reliable enough.The value of peritumoral characteristic and combination of radiomics and deep learning was also not explored.ObjectiveTo explore the value of CESM based radiomics and deep learning for the prediction of TNBC.MethodsIn this retrospective study,CESM images of 460 breast cancer（divided into training and testing sets with ratio of 4:1）patients confirmed by pathological examination were analyzed.ROIs（region of interests）of GTV（global tumor volume）,PERI3（peritumoral 3mm）,PERI6（peritumoral 6mm）,PERI9（peritumoral 9mm）,DILA3（dilated 3mm）,DILA6（dilated 6mm）.and DILA9（dilated 9mm）were segmented in CC（cranial caudal）,MLO（mediolateral oblique）of low-energy images and subtracted images.Radiomics features were extracted from ROIs of GTV,PERI3,PERI6,and PERI9 in respective images and normalized by z-score normalization.Radiomics features in training set was selected by variance thresholding method,student t test or Wilcoxon rank sum test,and LASSO（least absolute shrinkage and selection operator）with 10-fold cross validation.GTV_MLO,GTV_CC,GTV_CON,GTV_REC models were constructed based on radiomics features in GTV of MLO,CC,low-energy,and subtracted images,respectively.GTV_ALL,COM₃,COM₆,and COM₉ models were constructed based on radiomics features of GTV,combination of radiomics features of GTV and PERI3,GTV and PERI6,GTV and PERI9 in training set,respectively.Transform learning based on Resnet34 model was used in deep learning analysis.Deep learning image based models of DL_GTV₁,DL_DILA3₁,DL_DILA6₁,and DL_DILA9₁ were built using images of ROI of GTV,DILA3,DILA6,and DILA9 in training set,respectively.Deep learning patient based models of DL_GTV₂,DL_DILA3₂,DL_DILA6₂,and DL_DILA9₂ were built using images of ROI of GTV,DILA3,DILA6,and DILA9 in training set,respectively.Combination of best radiomics with deep learning models named DL-RAD was also built.All models were evaluated by ROC（receiver operating characteristic）curve analysis,accuracy,specificity,sensitivity,precision.F1-score,balanced accuracy and DCA（decision curve analysis）.AUCs（area under ROC curves）of different models were compared by Delong test.Results1.Model COM₉ performed best in 8 radiomics models with the highest AUC of 0.87 and the highest accuracy of 0.81.The AUC of model COM₉ was higher than AUC of model GTV_ALL,GTV_MLO,GTV_CC,GTV_CON,and GTV_REC significantly.2.Model GTV_ALL performed better than model GTV_MLO,GTV_CC,GTV_CON,and GTV_REC.3.In 8 deep learning models,model DL_DILA9₂ performed better than other models with the highest AUC of 0.86 and the highest accuracy of 0.84.Patient based deep learning models performed better than respective image based deep learning models.4.Model DL-RAD was constructed based on the combination of model COM₉ with model DL_DILA9₂.Model DL-RAD performed better than model COM₉ and model DL_DILA9₂ with the higher AUC of 0.91 and higher accuracy of 0.87.Model DL-RAD performed best in all the models in this study.The AUC of model DL-RAD was higher than the AUC of model GTV_ALL,GTV_REC,GTV_CC,GTV_MLO,GTV_CON,COM₃,COM₉,DL_GTV₁,DL_GTV₂,DL_DILA3₁,DL_DILA3₂,DL_DILA6₁,DL_DILA6₂,and DL_DILA9₁（P<0.05）.In DCA,net benefit was higher than model "all positive" and "all negative" when the diagnostic threshold was between 0.10 and 0.95.Conclusions1.CESM based radiomics and deep learning models were valuable for the prediction of TNBC.2.Model based on all intratumoral radiomics features performed better than models based on part of intratumoral radiomics features.3.Peritumoral information in CESM images had additive value for the prediction of TNBC.4.The combination of radiomics and deep learning improved prediction ability than radiomics or deep learning alone.Part ⅡThe Value of intratumoral and peritumoral CESM radiomics for the prediction of histological grade and Ki67 of triple-negative breast cancerBackgroundHistological grade and expression status of Ki67 of breast cancer were valuable factors for diagnosis and treatment,especially for triple-negative breast cancer（TNBC）.TNBC with higher histological grade and expression level of Ki67 prone to be more aggressive.Preoperative diagnosis method of histological grade and expression level of Ki67 of breast cancer is immunohistochemistry based on biopsy.Given limited tumor tissue,that method can not characterize the tumor comprehensively,accurately,and dynamically.The histological grade of breast cancer is possibly be underrated.The result of Ki67 expression level in immunohistochemical examination is easily affected by many factors such as specimen processing and staining.Preoperative diagnosis of histological grade and expression level of Ki67 of TNBC comprehensively,accurately using noninvasive method is meaningful.Radiomics features are extracted from medical images.It makes the diagnosis of histological grade and expression level of Ki67 of breast cancer noninvasively possible.As a novel medical imaging examination method,CESM（contrast-enhanced spectral mammography）was widely used in radiomics studies.However,value of CESM based radiomics for the prediction of histological grade and expression status of Ki67 of TNBC is not be explored.It is of clinical significance to explore the value of CESM based radiomics for prediction histological grade and expression status of Ki67 of TNBC.ObjectiveTo explore the value of intratumoral and peritumoral CESM radiomics for the prediction of histological grade and expression status of Ki67 of TNBC.MethodsCESM images of 150（142,150 for the prediction of histological grade and expression status of Ki67,respectively）TNBC patients confirmed by pathological examination were retrospectively analyzed.Patients for the prediction of histological grade and expression status of Ki67 were divided into training and testing sets with ratio of 3:1,respectively.ROIs（region of interests）of GTV（global tumor volume）.PERI3（peritumoral 3mm）,PERI6（peritumoral 6mm）,and PERI9（peritumoral 9mm）were segmented in CC（cranial caudal）,MLO（mediolateral oblique）of low-energy images and subtracted images.Radiomics features were extracted from all ROIs in respective images and normalized by z-score normalization.Radiomics features in training set was selected by variance thresholding method,student t test or Wilcoxon rank sum test,and LASSO（least absolute shrinkage and selection operator）with 10-fold cross validation.For the prediction of histological grade,GTV_HIS,COM3_HIS,COM6_HIS,and COM9_HIS models were constructed based on radiomics features of GTV,combination of radiomics features of GTV and PERI3,GTV and PERI6,GTV and PERI9 in training set,respectively.For the prediction of expression status of Ki67,GTV_KI67,COM3_KI67,COM6_KI67,and COM9_KI67 models were constructed based on radiomics features of GTV,combination of radiomics features of GTV and PERI3,GTV and PERI6,GTV and PERI9 in training set,respectively.All models were evaluated by ROC（receiver operating characteristic）curve analysis,accuracy,specificity,sensitivity,precision,F1-score,balanced accuracy.AUCs（area under ROC curves）of different models were compared by Delong test.Results1.For the prediction of histological grade,the AUC of model GTV_HIS,COM3_HIS,COM6_HIS,and COM9_HIS was 0.77,0.78,0.78,and 0.78,respectively,while the accuracy was 0.61,0.67,0.78,0.75,respectively.The AUC of model COM3_HIS,COM6_HIS,and COM9_HIS was higher than that of model GTV_HIS（P>0.05）.The evaluation metrics of COM6_HIS was higher or not lower than respective evaluation metrics of other models.2.For the prediction of expression status of Ki67,the AUC of model COM3_KI67,COM6_KI67,and COM9_KI67（0.79,0.71,0.73,respectively）was higher than that of model GTV_KI67（0.59,P>0.05）.Furthermore,the accuracy of model COM3_KI67,COM6_KI67,and COM9_KI67（0.76,0.68,0.79,respectively）was also higher than accuracy of model GTV_KI67（0.63,P>0.05）.Model COM9_KI67 acquired highest sensitivity of 0.81 while the model COM6_KI67 get the highest specificity of 0.83.ConclusionsCESM radiomics was valuable for the prediction of histological grade and expression status of Ki67 of TNBC.Peritumoral radiomics features improved the prediction ability of histological grade and expression status of Ki67 of TNBC.