Abstract:

Objective To investigate the value of intratumoral and peritumoral ultrasound radiomics features in predicting programmed death-ligand 1 （PD-L1） expression status in breast cancer and to construct an interpretable machine learning model. Methods A total of 101 breast cancer patients admitted at Nanjing Drum Tower Hospital from July 2017 to May 2022 were selected as the study subjects， including 48 cases of PD-L1-positive and 53 cases of PD-L1-negative. Patients were randomly divided into a training set （n=71） and a validation set （n=30） at a 7∶3 ratio. The age， tumor location， maximum tumor diameter， minimum tumor diameter， breast imaging reporting and data system （BI-RADS） category， histological grade and molecular subtype were compared between patients with PD-L1 positive and PD-L1 negative. A total of 976 radiomics features were extracted from both intratumoral and peritumoral 5 mm regions. Optimal radiomics features were selected using t-test and least absolute shrinkage and selection operator （LASSO） regression， and a radiomics score （RS） was constructed. The extreme gradient boosting （XGBoost） machine learning algorithm was used to construct a combined model for predicting PD-L1 expression. Model performance was evaluated using receiver operator characteristic （ROC） curve analysis. Interpretability of the XGBoost model was assessed using Shapley additive explanation （SHAP） analysis. Results In the training set， there was a statistically significant difference in histological grade between PD-L1-positive （n=33） and PD-L1-negative （n=38） groups （Z=-2.56， P=0.010）， whereas there were no statistically significant differences in age， tumor location， maximum tumor diameter， minimum tumor diameter， BI-RADS category， or molecular subtype （all P>0.05）. A total of 9 and 8 optimal radiomics features were screened from the intratumoral and peritumoral regions， respectively. Based on the above features， the intratumoral RS and peritumoral RS were respectively constructed. The intratumoral RS=-0.514 + 0.428×original_ ngtdm_ Contrast-0.153×wavelet_ HL_ glcm_ Correlation-1.224×wavelet_ HH_ glcm_ Id + 0.376×exponential_ gldm_ SmallDependenceLowGrayLevelEmphasis + 0.034×exponential_ glcm_ Idmn-1.006×gradient_ ngtdm_ Busyness + 0.365×gradient_ ngtdm_ Coarseness-0.893×logarithm_ glcm_ DifferenceVariance + 0.291×square_ glcm_ DifferenceVariance. The peritumoral RS=-0.463 + 0.391×wavelet_ LH_ firstorder_ Mean-0.715×wavelet_ LH_ firstorder_ Skewness-0.421×wavelet_ LH_ glcm_ Correlation-0.444×wavelet_ LL_ gldm_ DependenceNonUniformityNormalized + 0.859×logarithm_ glcm_ Correlation + 0.835×logarithm_ glcm_ InverseVariance + 0.392×square_ glcm_ Imc2-0.463×square_ glrlm_ GrayLevelNonUniformityNormalized. A combined model incorporating intratumoral RS， peritumoral RS， and histological grade was built using the XGBoost model. ROC curve analysis demonstrated that， the area under the curve of the combined model for predicting PD-L1 expression status in breast cancer patients was 0.909 （95%CI：0.817-0.965） and 0.864 （95%CI：0.690-0.961） in the training and validation sets， respectively. SHAP analysis revealed that peritumoral RS was the most predictive feature in the XGBoost model， followed by intratumoral RS and histological grade. Conclusions The XGBoost model integrating ultrasound intratumoral RS， peritumoral RS， and histological grade effectively predicts PD-L1 expression status in breast cancer， with peritumoral RS being the most valuable predictive feature.

Key words: Breast neoplasms, Ultrasonography, mammary, Forecasting, B7-H1 antigen