As reported in cancer statistics in 2019, breast cancer (BC) is one of the most common malignancies in women worldwide, ranking first in prevalence and second in mortality [1]. Histological type and grade, proliferative activity and status of human epidermal growth factor receptor 2 (HER2) and hormone receptors expression are increasingly accepted as the most common characteristics influencing the choice of treatment options, the evaluation of therapy responses and the prediction of survival for patients with BC [2, 3]. Consistently, BC is categorized into four different molecular subtypes based on expression of hormonal receptors and HER2 status, including Luminal subtype: hormone receptor positive, HER2 negative; Luminal/HER2 positive subtype: hormone receptor positive, HER2 overexpression or amplified; HER2 positive non-luminal subtype: hormone receptor negative, HER2 overexpression or amplified; Triple-negative (TN) subtype: hormone receptor negative and HER2 negative [4]. However, this molecular subtype classification is mainly achieved through invasive procedures, such as biopsy-based immunohistochemical (IHC) staining or fluorescence in situ hybridization (FISH) method [5]. Moreover, heterogeneity existed in different BC lesions from individual BC patients, even in a same lesion from a single BC patient [6]. Therefore, the local biopsy could not represent the whole tumor lesion or the distant unexamined metastatic lesions due to the existed heterogeneity in BC lesions. A reliable and non-invasive imaging methodology is needed to identify these gene expression profiles to develop personalized treatment and improve early prognosis prediction for patients with BC [7, 8, 9, 10].

As a hybrid imaging allowing for simultaneous anatomic imaging and molecular functional imaging,${}^{18}$F-Fluorodeoxyglucose positron emission tomography/computed tomography (${}^{18}$F-FDG PET/CT) imaging is already widely used in tumor detection, tumor staging, treatment response evaluation and survival prediction based on the provided functional information about tumor glucose metabolism for a variety of tumors, including BC [11]. Standardized uptake value (SUV), metabolic tumor volume (MTV) and total lesion glycolysis (TLG) are most commonly used semi-quantitative parameters extracted from PET images [12]. Furthermore, these traditional PET parameters have been found to be correlated to hormone receptors status and molecularly defined subtypes for patients with BC [13, 14]. Nevertheless, especially for SUV as a single pixel value, it is not able to adequately reflect the glucose metabolism of the whole tumor with heterogeneity [15].

It is now recognized that standard medical images may contain more useful information than that we can see by using our naked eyes. Radiomics is an emerging hot topic in medical imaging, which is actually a high-throughput extraction of quantitative metrics from medical images via computational post-processing techniques [4, 9, 10]. Indeed, radiomics is based on the assumption that the extracted heterogeneity descriptors from medical images are linked to genotypic and molecular characteristics of the tumor lesions, and is believed to be a promising model to guide patient management in personalized medicine [16]. To date, most studies about radiomics in relation to molecular characterization of BC are ultrasound [7], mammographic [17] and magnetic resonance imaging (MRI) radiomics [9, 10, 18, 19, 20]. Few data regarding to PET/CT radiomics for BC patients are available, especially for the performance of radiomics models based on PET/CT images in molecular subtype determination [4, 21, 22].

In the present investigation, the heterogeneity in ${}^{18}$F-FDG PET/CT images of primary BC lesions was quantitatively measured by radiomic analyses and several multi-feature radiomic signature models were developed to evaluate the predictive performance of radiomics based on PET/CT images in molecular subtype determination for patients with BC.

Results with respect to correlation between molecular subtypes of BC and imaging-derived parameters had been published previously [12, 13, 14, 15]. Especially, texture feature quantification and analyses based on medical images are of increasing interest to a large number of investigators addressing the interaction between quantitative imaging radiomic features and tumor molecular profiling [4, 16]. Moreover, considering breast cancer is a heterogeneous class of tumor, radiomic analyses aiming at elucidating the tumor heterogeneity reflected in medical images would enable an accurate prediction of tumor biological characteristics and genetic profiling. The primary objective of this current study was to evaluate the potential of ${}^{18}$F-FDG PET/CT derived radiomic signature models to distinguish between different molecular subtypes of BC, because of the unique advantage of ${}^{18}$F-FDG PET/CT imaging over other imaging modalities, such as mammography [17], ultrasound [7] and MRI [9]. As shown, BC lesions with luminal subtype exhibited significantly lower FDG radioactivity uptake (SUVmax, SUVmean and SUVpeak) compared to that with non-luminal subtype. Similarly, all of the three SUV parameters were found to be markedly underrepresented in BC patients with non-TN subtype in contrast with BC patients with TN subtype. Whereas, no substantial association was demonstrated between HER2+ BC patients and HER2– BC patients for all the five conventional PET parameters. These results demonstrated that different molecular subtypes of BC underlined the differences described in FDG uptake by BC lesions in vivo, which is increasingly recognized because it was previously reported that BC lesions with an aggressive subtype, such as non-luminal or TN, tended to be with an increase in glucose metabolism which could be captured and reflected in SUV parameters (SUVmax, SUVmean and SUVpeak) [13, 14, 15]. Whereas, PET volumetric parameters, such as MTV and TLG, which are always used to represent the tumor burden, were found to be not significantly different between different BC subgroups with different molecular subtypes. With regard to the discrimination between HER2+ and HER2– subgroups by conventional PET parameters, because both lower-glucose-metabolic BC with luminal subtype and higher-glucose-metabolic BC with TN were categorized into HER2– subgroup, which would lead to the negative results about HER2+ vs. HER2– classification by using conventional PET parameters.

Expectedly, established multivariate radiomic signature models as shown in Supplementary Fig. 1 remarkably outperformed all the significant SUV parameters in discrimination between different molecular subtypes of BC. In comparison with previous literature with regard to the association between radiomic features based on PET images alone and molecular subtypes of BC [4, 21, 22], our current investigation had a larger sample size and performed a comprehensive radiomic analysis involving both PET-based and CT based radiomic features. Particularly in contrast with results from ZY Yang et al. [21], which only focused on the discrimination between ER+ and ER–, our study consist of analysis with regard to Luminal vs. Non-Luminal, HER2+ vs. HER2– and TN vs. non-TN. In this context, the results listed in our investigation would provide an essentially complementary information about PET/CT derived radiomic analyses in molecular characterization of BC. Moreover, because selecting individual radiographic parameters which were not remarkably correlated to each other could provide added information but not overlapping messages with respect to various aspects of tumor lesion heterogeneity, multivariate radiomic signature model as presented in Supplementary Fig. 1 but not individual radiomic feature was used in our study to directly evaluate the performance of PET/CT derived radiomics in molecular subtype classification of BC, which was lacking in previous relevant investigations. To some extent, molecular subtype classification is a brilliant method to reflect the heterogeneity in BC and is helpful for personalized medicine of BC patients [2, 3]. For BC patients with HER2 enrichment, a sophisticated intracellular signaling initiated by HER2 activation on the membrane of breast cancer cells leads to an enhanced metabolism and a subsequently aggressive behavior of HER2-enriched breast cancer in contrast with breast cancer without HER2 accumulation [23]. Considering the tumor-promoting action of HER2 in breast cancer, targeted therapies against HER2 were developed and applied to clinical management, namely trastuzumab (Trade name: Herceptin) treatment [24]. To improve the performance of HER2-targeted therapy for BC in clinic and make a dramatic progress in precision medicine, a precise assessment of the target HER2 prior to treatment is essential. As known, HER2 status was most commonly tested via IHC examination, and cases with ambiguous results usually refer to FISH assay. A subgroup radiomic analysis focusing on a cohort of BC patients who underwent a FISH assay to finally identify HER2 status because of their equivocal IHC results of HER2 status is planed in our future study. The abilities of PET/CT derived radiomic signature models to discriminate molecular subtypes of BC suggest an underlined molecular profile-involving mechanism for heterogeneity reflected by PET/CT derived radiomic features [6]. Even though BC patients could benefit from current molecular subtype classification to guide their clinical management, such as treatment options, outcome prediction and survival prognosis, more and more investigation have discovered heterogeneity existed in the same molecular subtype of BC within individuals and between individuals. The existed heterogeneity in sensitivities to some targeted therapies and neoadjuvant treatment for BC patients with the same molecular subtype and the development of acquired treatment resistance prompt us to further determine the performance of radiomic signature model based on PET/CT images in these concerning issues, which is what we strive to do in near future.

The ultimate goal of accurate molecular subtype classification for BC patients is to individualise therapy and achieve precision medicine. Multidisciplinary therapeutic approaches, including hormone therapy, HER2-targeted therapy, chemotherapy and radiotherapy are precisely chosen for BC patients based on molecular profile determination. Radiomics and artificial intelligence development have not only shed light on molecular subtype profiling, but also for other various aspect of clinical practice for BC patients, such as diagnosis, treatment response prediction and survival prognosis [25]. Especially for neoadjuvant radiotherapy or chemoradiotherapy, radiomics and artificial intelligence have exhibited a potential for clinical translation in tailed treatment prediction [26, 27, 28]. With respect to BC patients, radiological image derived artificial intelligence-based decision supporting system for radiotherapy would be beneficial to patient consultation, target volume delineation, automated treatment planning, treatment delivery and response evaluation and prediction [29]. Furthermore, an enhanced prediction of neoadjuvant chemoradiotherapy response by using PET/CT radiomic models with an addition of HER2 information suggested that accurate HER2 status determination based on radiomics would be remarkably helpful to comprehensive treatment strategy involving radiotherapy and targeted therapy [28].

Despite those encouraging results depicted in this study, several limitations existed in this current work. First, given the relatively small sample size in this radiomic analysis, further investigation involving a larger number of cases is warranted to confirm the results obtained in the study. Second, the varieties in acquisition, reconstruction and delineation settings maybe affect the repeatability of radiomic analyses due to the retrospective nature of this study. A prospective, multicenter research with a standardization and optimization for these settings is needed to overcome these drawbacks. Then, a relatively limited spatial resolution of PET may result in an unsatisfactory performance of PET related radiomics in breast cancer, in particular for patients with small lesion. In addition, this study only focused on primary breast cancer lesions but not on metastatic lesions, therefore a further investigation involving both primary and metastatic lesions is required to complete this study [30]. Last but not least, to comprehensively determine the phenotype-genotype interaction, radiomic analyses based on not only a binary-classification but also on a multi-classification for molecular subtype clustering of BC should be included, which is exactly what we aim to do in future.

Taken together, distinct imaging phenotypes driven by different molecular profiling of BC may be captured in ${}^{18}$F-FDG PET/CT images and quantitatively measured by radiomic features. Established multivariate radiomic signature models based on ${}^{18}$F-FDG PET/CT images could provide more information than conventional individual PET parameters which are currently widely used for clinical imaging, enabling a precise reflection of tumor heterogeneity.

JL, HB and YZ contributed equally and are co-first authors; JL and HB designed experiments; JL, HB and YZ performed experiments; JL, HB, YZ and XL wrote the manuscript; GY, YG and ZW analyzed data; YG synthesized and interpreted the results. XL, WM and WX supervised the study and made manuscript revisions. All authors gave final approval of the version to be published and have agreed on the submitted journal. And all authors agreed to be accountable for all aspects of the work.

All procedures performed in studies involving human participants were in accordance with the declaration of Helsinki and relevant approved guidelines and approved by the Ethical Review Committee of Tianjin Medical University Cancer Institute and Hospital, and the ethical approval code is Ek2018125. Written informed consent from all patients was waived in the present investigation.

We would like to thank all the study participants, research staff, and students who participated in this work. In addition, we are really appreciated for the invaluable comments on this study from all the reviewers.

This work was supported by grants from the National Natural Science Foundation of China (grant nos. 81801781, 82072004, 2018ZX09201015), the Tianjin Science and Technology Committee Fund (grant nos. 18PTZWHZ00100 and H2018206600), the Science & Technology Development Fund of Tianjin Education Commission for Higher Education (grant nos. 2018KJ057 and 2018KJ061).

The authors declare no conflict of interest.