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

Breast cancer (BC) has already become the most frequently diagnosed tumor and the leading cause of cancer-related death among women worldwide [1, 2]. According to the different expressions of specific molecular receptors, BC can be classified into various subtypes, which are tightly associated with the treatment strategy selection and prognosis [3, 4]. However, even with classic targeted therapies, a certain proportion of patients are insensitive to treatments and develop varying degrees of resistance [5‐8]. Therefore, more attention should be paid to new molecular targets to assist in the diagnosis and treatment of BC.

Most recent studies have investigated the roles of androgen receptor (AR) in BC patients, showing that AR has different subtypes with considerable levels of expression [9, 10]. AR and estrogen receptor (ER) competitively bind to estrogen-dependent signaling pathways, and this crosstalk may contribute to acquired resistance to endocrine therapy [11, 12]. Besides, the expression of AR is highly correlated with the amplification of human epidermal growth factor 2 (HER2), which is coupled with AR activation and enhances the oncogenic function of certain signaling pathways [13, 14]. Triple-negative BC (TNBC) lacks recognized molecular targets, in which the expression of AR is related to ER-regulated gene transcription and a lower rate of pathological complete response (pCR) to neoadjuvant chemotherapy [15, 16]. Consequently, early recognition of AR expression not only helps predict prognosis but also guides the application of AR-targeted therapy, effectively improving patients’ sensitivity to traditional therapies and prolonging their overall survival.

At present, the expressions of molecular receptors are generally detected through immunohistochemistry (IHC). Such an approach is mainly based on biopsy as an invasive procedure, and it is rarely repeatable and unrepresentative due to small tissue samples and intratumor heterogeneity [17]. ¹⁸F-FDG PET/CT (Fluorine-18 fluorodeoxyglucose positron emission tomography/computed tomography), which combines functional metabolic quantification and morphological imaging, is commonly used for initial staging, post-treatment assessment and re-staging in BC [18, 19]. Radiomics, which uses medical imaging to noninvasively quantify intratumoral heterogeneity, has emerged as a translational research topic in BC [20]. Recent literature has indicated that radiomic features obtained from ¹⁸F-FDG PET/CT can contribute to the prediction of tumor molecular subtypes, while there are still limited studies focusing on the role of AR as a predictive parameter [17].

In the present study, we aimed to develop a machine-learning model for individual prediction of AR expression in BC patients by exploring the correlations between radiomic features, clinicopathological factors, and AR status. In addition, our findings might help identify the patients who could benefit from the AR-targeted treatment.

Some studies have focused on the correlation between the status of other receptors and the extensive expression of AR, indicating that it can promote the proliferation of cancer cells [24]. Accumulating preclinical and clinical evidence supports that some patients can benefit from AR inhibitor therapy, which especially improves the prognosis of TNBC patients [25‐27]. In order to avoid the limitations of IHC, recent studies have demonstrated the feasibility of radiomics combined with clinical data for predicting intrinsic molecular subtypes and the expressions of receptors in BC [28, 29]. However, only very few studies have focused on predicting the expression of AR status. In our present study, MTV, SHAPE_Sphericity_CT, GLCM_Contrast_CT extracted from ¹⁸F-FDG PET/CT could predict the AR status in BC patients. We further developed a machine-learning model with a higher AUC score on the ROC curve (0.832, 95% CI: 0.697–0.924).

Previous studies have shown that standard parameters (SUV_max, SUV_mean, and MTV) extracted from PET images are associated with ER or PR status [17, 30, 31]. In our univariate analysis, there was no obvious correlation between SUV_max, SUV_mean, and AR status. Huang et al. have concluded that ER and PR status have significant differences between AR + and AR- groups [23]. ER and PR status were weakly correlated with the expression of AR (p = 0.061, p = 0.127) in our study. The reason was attributed to the residual confounding factors, and it needs to be further investigated in a larger cohort population [32, 33]. However, high MTV showed a tendency to be associated with AR negativity (p = 0.032). Kaida et al. have also found a similar situation, in which a feeble correlation exists between high MTV and ER-negative status [31]. In further textural analysis, SHAPE_Sphericity_CT (p = 0.004), GLCM_Contrast_CT (p = 0.008), NGLDM_Coarseness_CT (p = 0.008), SHAPE_Sphericity_PET (p = 0.012), and NGLDM_Coarseness_PET (p = 0.004) selected by LASSO and t-test were correlated with the expression of AR. In multivariable logistic analysis, only SHAPE_Sphericity_CT and GLCM_Contrast_CT were included in the prediction model. Due to the limited density resolution of the PET images, they performed worse than CT images in screening and extracting meaningful radiomics signatures [34]. Currently, GLCM (grey-level co-occurrence matrix) textural features are used to classify the tumor grade of BC [35‐38]. Here, we are the first to report the independent performance of GLCM_Contrast_CT in predicting AR molecular markers (p = 0.018).

Based on the forward stepwise logistic regression, we established a combined model, in which the accuracy was usually assessed by discrimination, calibration, and clinical applicability [39, 40]. According to its AUC (AUC = 0.832), this model presented a better efficacy in the discrimination of AR status. Cross-validation and Bootstrap have proved to be used for the predictive performance of a small sample [41]. Furthermore, using the above methods for internal validation, we finally obtained a very stable value of AUC (AUC = 0.833). Some studies have proved that the c-index, with its 95% CI, can provide a more comprehensive assessment of calibration [42, 43]. Combined with the p-value of the H-L test, our model showed a higher capacity for stability [44]. DCA can be used to map the utility of a net benefit assessment model in decision-making within a clinically reasonable risk threshold [45]. The DCA curve of this study visualized that the diagnostic model achieved better net benefit than the all-intervention or no-intervention strategy within a larger threshold range, further proving the excellent clinical applicability of this model.

To the best of our knowledge, we originally established a combined model to predict AR expression based on ¹⁸F-FDG PET/CT imaging. There were some limitations in our present study due to the single-center and retrospective design. Next, due to the novelty of the molecular receptor detection of AR, the number of patients included in this study was relatively small. Thirdly, except for internal validation, the resulting model needed to be validated with an external cohort to improve the confidence of clinical applicability. Lastly, a classical supervised classification algorithm should be used to construct a prediction model, and deep learning-based features from PET/CT images need to be verified in further studies.

In conclusion, by combining clinicopathological factors and radiomic features extracted from ¹⁸F-FDG PET/CT images, we established a model to predict AR expression in BC patients with a single primary lesion. Our model could serve as a novel strategy to select patients who could benefit from anti-AR treatment and to assist clinicians in making clinical decisions.