The Application Of DBT-based Radiomics In Predicting Luminal And Non Luminal Breast Cancer

Objective:1.Three models were constructed based on different body position views in Digital Breast Tomosynthesis(DBT)examination to evaluate the value of DBT image based radiomics technology in preoperative assessment of whether invasive breast cancer is Luminal classification.2.A nomogram constructed by combining the Radiomics Score(Rad-score)with clinical independent risk factors,to explore the value of this nomogram in the preoperative evaluation of whether Luminal classification is applied to patients with invasive breast cancer.Methods:1.The standard images and clinicopathological data of patients with invasive breast cancer confirmed by pathology in our hospital from January 2020 to September 2022 were retrospectively collected,and 217 cases were finally included,including 167 cases of Luminal type and 50 cases of non Luminal type.The patients were randomly divided into training set(n=151)and validation set(n=66)according to the ratio of 7:3.2.Manually delineate several layers of Region of Interest(ROI)on the DBT image to further generate a Volume of Interest(VOI).By extracting and filtering the DBT imaging characteristics of Craniocaudal(CC)and Medial Oblique(MLO)positions,the logical regression(LR)algorithm is used to establish the CC bit model,MLO bit model,and CC+MLO bit model,respectively.3.According to the clinicopathological data of patients,correlation statistical analysis was performed on age,lesion location,maximum diameter of the lesion,presence or absence of calcification,breast imaging reporting and data system(BI-RADS)typing,Ki-67 expression level,and presence or absence of axillary lymph node metastasis.The clinicopathological characteristics with P<0.05 were then determined as clinical independent predictors using binary logistic regression method.4.The CC+MLO bit radiomics model is combined with clinical independent predictors to establish a joint clinical model,and the model is used to draw a nomogram to visualize the model.The diagnostic effectiveness of the model was evaluated by calculating the area under curve(AUC)using the receiver operating characteristic(ROC)curve,sensitivity,specificity,and accuracy.The consistency between the predicted value and the actual value was evaluated using a calibration curve.The decision curve analysis(DCA)was used to quantify the net benefit assessment of imaging omics models and nomograms under different threshold probabilities,and compare the clinical applicability of each model.Results:The CC bit and MLO bit DBT images extract 1792 radiomics features from each posture,and screen out 7 key radiomics features each.These 14 key radiomics features are statistically significant(P<0.05).The 14 radiomics features are respectively generated into Rad-scores for their respective posture models through linear fitting,and incorporated into the LR algorithm to establish their respective radiomics models.A horizontal comparison of the LR models constructed separately found that the LR prediction model constructed using a dual perspective approach based on the CC+MLO bit had an AUC(95% CI)of 0.853(0.798,0.907),a sensitivity of 0.810,a specificity of0.871,and an accuracy of 0.801 in the training set.In the validation set,the indicators were 0.846(0.740,0.934),0.863,0.833,and 0.803,respectively,and their overall performance was better than the two LR prediction models constructed based only on the CC bit or MLO bit single perspective.After relevant statistical analysis,Ki-67 status and BI-RADS typing were selected as clinical independent predictors.Together with CC+MLO-Rad-score,the multifactor LR algorithm was incorporated and a joint clinical model was constructed.The AUC(95% CI)of the model in the training set was 0.873(0.820,0.919),sensitivity was 0.838,specificity was 0.943,and accuracy was 0.809.In the validation set,each index was 0.865(0.769,0.948),0.849,0.867,and 0.821,respectively,The efficacy of this combined clinical model is higher than that of CC,MLO,and CC+MLO models.Therefore,the combined clinical model is used as the optimal model to construct a nomograph.The calibration curve shows a good consistency between the evaluation of the nomogram and the actual molecular typing status of Luminal.The DCA curve shows a threshold probability range of 0.3-0.7,and the clinical net benefit of the nomogram model is higher than that of the pure radiomics model.Conclusion:1.The efficiency of the radiomics model based on CC+MLO dual position in predicting Luminal and non Luminal breast cancer is better than that of CC and MLO single position radiomics model,which can provide certain reference value for subsequent radiomics research.2.The routine clinicopathological factors Ki-67 status and BI-RADS classification,as independent clinical predictors,will help enrich the radiomics model and further help to differentiate the molecular classification of breast cancer.3.The combined clinical model built by the combination of breast cancer DBT characteristics and clinical independent predictors is the optimal model.The nomogram built by this model has good predictive performance.The DCA curve shows that within the threshold probability range of 0.3-0.7,the clinical net income of the nomogram model is higher than that of the simple icomics model,which can directly,quickly and noninvasive predict whether breast cancer patients are classified into Luminal molecular types,to provide effective basis and guidance for further clinical diagnosis and treatment and judgment of prognosis.