Assessment of androgen receptor expression in breast cancer patients using 18 F-FDG PET/CT radiomics and clinicopathological characteristics

Abstract

Objective: In the present study, we mainly aimed to predict the expression of androgen receptor (AR) in breast cancer (BC) patients by combing radiomic features and clinicopathological factors in a non-invasive machine learning way.

Materials and methods: A total of 48 BC patients, who were initially diagnosed by ¹⁸F-FDG PET/CT, were retrospectively enrolled in this study. LIFEx software was used to extract radiomic features based on PET and CT data. The most useful predictive features were selected by the LASSO (least absolute shrinkage and selection operator) regression and t-test. Radiomic signatures and clinicopathologic characteristics were incorporated to develop a prediction model using multivariable logistic regression analysis. The receiver operating characteristic (ROC) curve, Hosmer-Lemeshow (H-L) test, and decision curve analysis (DCA) were conducted to assess the predictive efficiency of the model.

Results: In the univariate analysis, the metabolic tumor volume (MTV) was significantly correlated with the expression of AR in BC patients (p < 0.05). However, there only existed feeble correlations between estrogen receptor (ER), progesterone receptor (PR), and AR status (p = 0.127, p = 0.061, respectively). Based on the binary logistic regression method, MTV, SHAPE_Sphericity_CT (CT Sphericity from SHAPE), and GLCM_Contrast_CT (CT Contrast from grey-level co-occurrence matrix) were included in the prediction model for AR expression. Among them, GLCM_Contrast_CT was an independent predictor of AR status (OR = 9.00, p = 0.018). The area under the curve (AUC) of ROC in this model was 0.832. The p-value of the H-L test was beyond 0.05.

Conclusions: A prediction model combining radiomic features and clinicopathological characteristics could be a promising approach to predict the expression of AR and noninvasively screen the BC patients who could benefit from anti-AR regimens.

Keywords: 18F-FDG PET/CT; Androgen receptor; Breast cancer; Clinicopathological; Machine learning; Radiomics.

Fig. 1

The workflow of this study. CT image (A), PET image (B), PET/CT fusion image (D), coronary image (E), sagittal image (F), and transaxial image (G) of ¹⁸F-FDG PET/CT display an example of VOI for extracting imaging features of BC. A patient with stage IV of BC underwent PET/CT image showing a metabolically active right breast lesion

Fig. 2

The Lamda of LASSO regression. The least absolute shrinkage and selection operator (LASSO) was conducted to select the radiomic features of CT and PET. Using 10-fold cross-validation, the suitable value of tuning parameter Lambda (λ) in LASSO regression was selected, and a vertical line was drawn here. MSE: mean squared error

Fig. 3

The Pearson correlation coefficient matrix heatmap in the prediction of AR expression. The darker color presents a higher correlation between the two factors

Fig. 4

The nomogram of risk factors predicting the expression of AR. The importance of each variable that established the diagnostic model was visualized as the points

Fig. 5

A typical case of the relationship between MTV and AR expression. Two patients were diagnosed with IDC (invasive ductal breast carcinoma) successively in a year. The former (A: PET/CT fusion image, B: The VOI of the lesion, C: The histogram of VOI) was confirmed as AR-positive (SUV_max: 8.00; MTV: 10.76 cm³). The latter (D: PET/CT fusion image, E: The VOI of the lesion, F: The histogram of VOI) was confirmed as AR negative (SUV_max: 9.12; MTV: 21.32 cm³). These lesions were similar on PET/CT images but showed significant differences in the histograms of the radiomic features

Fig. 6

The evaluation of the prediction model of AR expression in internal validation using Bootstrap. The ROC curve of the combined model consisted of radiomic features and clinical characters (A); the calibration curve (green) after correction of bias fluctuates around the ideal curve (red), which revealed the good accuracy between the actual probability and predicted probability (B); decision curve showed that the prediction model led to a higher net benefit than intervention or no-intervention of all patients among a large range of threshold (C)