Decoding Intratumoral Heterogeneity of Breast Cancer by Multiparametric In Vivo Imaging: A Translational Study

Beyond individual variances between different tumors (intertumoral heterogeneity), intratumoral heterogeneity has become a major focus in oncology (1–4). Defined by the appearance of different histologic, genetic, and molecular subpopulations within single tumors, intratumoral heterogeneity creates a new challenge in cancer classification and drives neoplastic progression and therapeutic resistance (5–7). To date, 50% to 80% of malignant breast cancers do not display sufficient characteristics of any special histologic type, but sometimes show high grades of intratumoral heterogeneity (1, 2) and are subsequently classified as “invasive breast carcinoma of no special type” (IBC-NST; Supplementary Table S1 contains a list of abbreviations; ref. 8).

Pathologists, therefore, analyze tissue sections from different regions of the tumor, reporting the highest observed grade (9). However, tumor classification and subsequent diagnosis in the clinic remain mostly based on small tissue samples that represent a single point of information rather than the status of the whole lesion.

Molecular imaging modalities, such as PET and MRI, have the potential to provide whole-lesion information noninvasively. PET shows great benefits in staging and therapy response evaluation of lung cancer and lymphoma (10–12) but has not been transferred into routine clinical practice for breast cancer patients.

Routinely performed PET imaging analyses of tumors are based on mean or hot spot values of tracer uptake, such as maximal standardized uptake value (SUV_max) or “metabolic active tumor region” evaluation (13, 14). Wahl and colleagues described the beneficial value of the “PET response criteria in solid tumors” (PERCIST) and discussed the great advantages of hot spot analysis procedures over the mean value analysis (15). However, all of these procedures neglect a substantial part of the information, as only a small fraction of the tumor is considered. More sophisticated approaches such as radiomics, extracting large numbers of quantitative imaging features for tissue characterization (16, 17), or other types of voxelwise analyses generating parametric maps of functional parameters (18, 19) are rarely used in clinical routine PET.

We hypothesized that hybrid PET/MRI can provide a clinically relevant tumor profiling and phenotyping approach of breast cancer. Thus, in this study, clinically applied PET and MRI methods using [¹⁸F]Fluorodeoxyglucose ([¹⁸F]FDG)-PET and diffusion-weighted (DW)-MRI were combined with a unique approach of voxelwise analysis, which identified different phenotypes within single tumors. To mimic the heterogeneous nature of breast cancer, polyoma virus middle T antigen transgenic [tg(PyV-mT)] mice, that develop malignant tumors, passing through distinct stages of tumor progression, from atypical hyperplasia to mammary intraepithelial neoplasia and invasive carcinoma (20, 21), were investigated. Furthermore, the clinical potential of the method was proven, including a small cohort of five IBC-NST patients. Our clustering approach provides a measure for the number of clusters and corresponding thresholds for the subsequent identification of tumor phenotypes and allows an easy translation to other oncological entities.

All experiments were performed with female tg(PyV-mT) mice on a C57BL/6 background. All animals were housed under standardized environmental conditions with free access to food and water. The tg(PyV-mT) mice developed an average of five palpable tumors, composed of a mixture of glandular and cystic dysplasias and invasive carcinomas (21, 22).

Preclinical in vivo studies were performed in a sequential PET/MRI setup using a dedicated 7 T small-animal scanner [ClinScan (7T-MRI); Bruker BioSpin GmbH] and a small-animal PET scanner [Inveon dedicated PET (DPET); Siemens Healthcare]. Five anesthetized animals (1–2% isoflurane evaporated in breathing air) at the age of 19 to 22 weeks, bearing together 26 tumors were measured on temperature-controlled animal beds. Apparent diffusion coefficient (ADC) MRI measurements using a half Fourier-acquired single shot turbo spin echo sequence [HASTE; repetition time (TR) 5,000 milliseconds, echo time (TE) 112 milliseconds, pixel size 0.21 × 0.21 mm² in plane, 1 mm slice thickness, b-values 150, 600, 1000 s/mm²] and sequential [¹⁸F]FDG-PET/MRI [intravenously injection of 10 ± 2 MBq of [¹⁸F]FDG; 50 minutes sleeping uptake, 10 minutes emission scan, 13 minutes transmission scan] were performed on two consecutive days. PET data were reconstructed using the ordered-subsets expectation maximization algorithm with 3D post-reconstruction (OSEM3D; Inveon Acquisition Workplace, Siemens Healthcare). The anatomical MRI was measured directly before the PET scans or the ADC measurements with a 3D T2-weighted turbo spin echo sequence (TSE; TR 2500 milliseconds, TE 202 milliseconds, voxel size 0.27 × 0.27 × 0.27 mm³).

Although not included into this report, four further PET/MRI measurements, investigating the PET tracers [⁶⁸Ga]NODAGA-RGD, [¹¹C]methionine, [¹¹C]choline, and [¹⁸F]FMISO were performed with the same animals previously within 1 week.

An additional group of four tg(PyV-mT) mice at the age of 20 weeks was measured in a sequential PET/MRI setup including anatomical and DW-MRI and [¹⁸F]FDG-PET. The mice were injected intravenously with 10 ± 2 MBq of [¹⁸F]FDG. During the uptake time of 60 minutes, MRI measurements were performed, directly followed by the PET scans.

All experiments were approved by the Regierungspraesidium Tuebingen.

Following the last in vivo measurement, the mice were sacrificed, and tumors were excised and prepared for hematoxylin and eosin (H&E) histology and/or autoradiography following standard procedures (Supplementary Materials and Methods).

Five patients at the age of 66 ± 10 years with biopsy-proven breast cancer lesions [estrogen receptor- and progesterone receptor-positive (ER⁺/PR⁺) IBC-NST, G₂–G₃] of ≥1.5 cm in size were examined using a 3T-PET/MR scanner (Biograph mMR; Siemens Healthcare). After 60 minutes of [¹⁸F]FDG uptake, the simultaneous PET/MRI acquisition was performed (258 ± 4 MBq, 20 minutes PET acquisition, 4.3 mm maximal achievable resolution; ref. 23). The PET data were reconstructed using an OSEM3D algorithm, applying an MR-based attenuation map from segmented MR images provided by the vendor. For the MRI measurements, a dedicated four-channel breast MRI coil (Siemens Healthcare), that was installed during all PET and MRI measurements, was used to acquire the following MRI sequences within 20 minutes: short-time inversion recovery (STIR), DW echo planar imaging (EPI, TR 11,000 milliseconds, TE 76 milliseconds, pixel size 1.8 × 1.8 mm² in plane, 4 mm slice thickness, b-values 50 and 800 s/mm²), contrast agent (Gadovist; Bayer Vital GmbH) enhanced (CE) dynamic imaging [T1 fast low-angle shot (FLASH) 3D] and post contrast agent T1 FLASH 3D fat saturated (TR 8.75 milliseconds, TE 4.21 milliseconds, voxel size 0.9 × 0.9 × 0.9 mm³). ADC maps were calculated from the DW EPI.

All patients gave written informed consent for the purpose of anonymized evaluation and publication of their data. All reported investigations were conducted in accordance with the Helsinki Declaration and with national regulations.

Postsurgery, whole slide histology of the tumors was performed and correlated to the imaging data. The excised tissues were cut in 0.5 cm slices for macroscopic analysis and fixed in formalin and embedded in paraffin. Tissue samples were taken for routine microscopic evaluation in 4 × 4 cm² sections. Subsequently, 3 μm H&E sections were prepared and digitized (Axioscope; Leica Microsystems).

Inveon Research Workplace (IRW; Siemens Healthcare) was used for in vivo image analysis. The corresponding preclinical PET and MR images were fused and coregistered using capillaries filled with the tracer, which were placed on the animal beds as markers for the PET and MRI data fusion. The simultaneously acquired clinical data were coregistered by the scanner software. Volumes of interest (VOI) covering the entire regions (whole breasts or lesions, respectively) were drawn on the basis of the MRI data (T2-weighted TSE for preclinical data; postcontrast agent T1 FLASH for the clinical data).

Beyond the tumor-to-muscle ratios (TMR, preclinical data) and standardized uptake values (SUV, clinical data), a new voxel-based approach was developed for [¹⁸F]FDG-PET and DW-MRI using a Gaussian mixture model (GMM) in MATLAB (MathWorks), fitting multiple Gaussian distributions to the collective voxel data (24). For every set of data, several different mixture models with up to seven Gaussian distributions were fitted and compared using the Akaike information criterion (AIC) and Bayesian information criterion (BIC), as implemented in the MATLAB package. The intersections formed by the descending part of one Gaussian distribution and the ascending part of the following distribution were used as thresholds to separate the subpopulations. Every Gaussian distribution within the thresholds defined a distinct biomarker population. To illustrate the results, histograms of the voxel values were calculated and overlaid with the GMM fitting results.

For ADC and PET data, parametric maps were computed, where each identified population was assigned to a specific color. Subsequently, for the sequentially acquired cohort multiparametric [¹⁸F]FDG/ADC maps were computed for the entire tumors.

To confirm the preclinical in vivo [¹⁸F]FDG results, the GMM approach was transferred to the autoradiography data of eight tumors. Following a background correction, TMR maps were computed and pixel values were extracted. The collective data set of all tumor pixels within the eight tumors was fitted. Using the determined thresholds, parametric autoradiography maps were computed.

First results from the tg(PyV-mT) studies revealed that entire lesions (n = 26 tumors, 53–769 mm³ in size) could only be detected on the basis of anatomical MRI, as intratumoral [¹⁸F]FDG uptake and the ADC were very heterogeneous. Because of the extensive intratumoral heterogeneity, phenotypic identification of whole lesions in terms of intertumoral comparison using single value analysis, such as mean and maximal value analyses, was not feasible, as both methods omitted important information (Supplementary Fig. S1).

To decode the heterogeneous tumor uptake pattern that is shown for the [¹⁸F]FDG uptake of four representative tumors (Fig. 1A), a voxel-based analysis approach for the multiparametric imaging procedure was developed. Three mice at the age of 22 weeks bearing 18 tumors were analyzed by a voxelwise calculation of the TMR for every tumor, representing the training data set that was collectively fitted using the GMM.

Evaluating the AIC and BIC, a mixture of four uptake populations was determined as the best fit for the [¹⁸F]FDG data (Fig. 1B). The intersections of the four distributions at TMR = 2.3, 5.1, and 9.6 represented the thresholds between the uptake populations (Fig. 1C). Parametric color maps of the tumors were created by applying these thresholds on the image data (Fig. 1E). Comparing to histology, these maps showed a correlation of the two populations with low [¹⁸F]FDG (blue, TMR < 5.1) uptake with cystic tumor regions, whereas solid tumor regions correlated very well with the population with the highest uptake (red, TMR > 9.6; Fig. 1D and E). However, not every Gaussian distribution directly correlated to a specific histologic pattern. Because of the limited spatial resolution of the PET scanner (1.6 mm maximal achievable resolution; ref. 25), the population with intermediate uptake, representing TMR values between 5.1 and 9.6, indicated mixed regions in which more than one phenotype was present within single voxels.

To confirm this phenotypic correlation, the identified thresholds from the training data set were applied to a test data set of eight tumors (two mice at the age of 19 weeks) without using the GMM analysis again and correlated with ex vivo autoradiography results (Fig. 2). Therefore, the TMR maps of the autoradiography data of these eight tumors were calculated and analyzed using the GMM algorithm. A total of five Gaussian distributions were determined as the best model ex vivo (Fig. 2A). The thresholds were calculated as TMR = 3.2, 7.7, 16.2, and 41.6 and applied to the autoradiography data to compute parametric color maps. The numerical difference in the TMR thresholds between the in vivo and autoradiography measurements may be attributed to different shrinking behaviors of the tissues during the autoradiography process, including the initial freezing and subsequent drying during the exposure time as well as partial volume effects (PVE) in PET. The 41.6 threshold was not considered to be a real threshold as the two populations with the highest uptake values overlapped significantly and were grouped together, as a secured assignment of individual voxels to either population was not possible. As the first two blue populations (TMR < 3.2 and TMR 3.2–7.7) mainly corresponded to the background and spillover at the edges of the tumors, the number of dedicated tumor populations was reduced to 2 (TMR 7.7–16.2 and 16.2 < TMR). This result was expected due to the better spatial resolution of the autoradiography plate (0.05 mm pixel diameter) compared with the PET data, which led to a reduction of the mixed population that was observed in vivo, when more than one phenotype was present in a single voxel. The correlation of the in vivo histograms of single tumors (Fig. 2B) and the ex vivo histograms, as well as single slices of the autoradiography (Fig. 2C) with the PET data (Fig. 2D), revealed a similar uptake pattern as shown for a representative tumor. The mixed areas in the PET data were mainly correlated with the population showing positive uptake by autoradiography, indicating a high amount of solid tissue in the mixed areas of the in vivo data that was verified by the corresponding H&E histology. The lower values of these regions in the in vivo data could be explained by an underestimation of the uptake due to PVE.

The GMM approach was also applied to the ADC data. All tumor voxels within the training data set were fitted (Fig. 3A). According to the defined fitting criteria, five ADC populations were identified as the best model (Fig. 3B). As the first and second population represented background populations (data not shown), these two populations could be neglected. The thresholds of the three remaining populations were determined as ADC = 0.6 × 10⁻³ mm²/s and ADC = 1.0 × 10⁻³ mm²/s. Mapping these populations to 3D parametric color maps and correlating these maps with the H&E histology showed a correlation of the high ADC values (ADC > 1.0 × 10⁻³ mm²/s) with the cystic hyperplastic regions within the lesions, whereas the low ADC population (ADC < 0.6 × 10⁻³ mm²/s) correlated with the solid regions. Similar to the [¹⁸F]FDG populations, a mixed ADC population was observed (ADC: 0.6–1.0 × 10⁻³ mm²/s), where more than one phenotype was present within single voxels.

To compare the parameters of [¹⁸F]FDG and ADC, both classifications were applied to each tumor voxel within the training data set, to investigate the tumors' composition according to both parameters (Fig. 3C). Although the population with high [¹⁸F]FDG uptake (TMR > 9.6) correlated with solid regions showing high metabolic activity, ADC maps clearly revealed hyperplastic cystic regions (ADC > 1.0 × 10⁻³ mm²/s), which could not be identified by the [¹⁸F]FDG uptake due to the low uptake values < 5.1 TMR. The combination of both methods, [¹⁸F]FDG and ADC PET/MRI, provided the opportunity to identify solid and cystic regions and thus, characterized the whole tumor burden within the mice without the need for tissue sections.

On the basis of these results, a multiparametric study was designed that enabled the direct correlation of the voxel data of both modalities for in vivo phenotypic clustering. Because of the sequential imaging setup, the ADC and [¹⁸F]FDG data of four mice at the age of 20 weeks could be directly fused for analysis. The thresholds of the training data set were directly transferred to this test data set without using the GMM fitting again.

Furthermore, the sequential setup allowed the automatized clustering of the tumors into phenotypic regions based on the coregistered ADC and [¹⁸F]FDG values in every voxel, providing phenotypic maps of the tumors (Fig. 4A).

As expected, fusion of [¹⁸F]FDG and ADC in the same tumors enabled the identification of solid and cystic regions within the tumors (Fig. 4A). Furthermore, beyond this classification, the solid regions could be differentiated into two distinct phenotypic patterns, solid acinar and very rare solid nodular regions. Solid acinar regions were characterized by [¹⁸F]FDG uptake values higher than 9.1 TMR; as shown previously, cystic and solid nodular regions showed low [¹⁸F]FDG uptake. Although the cystic areas showed high ADC values (ADC > 1.0 × 10⁻³ mm²/s), solid regions were characterized by low ADC values. Thus, the solid nodular regions were characterized by a double-negative signature (low ADC and low [¹⁸F]FDG uptake), enabling the further differentiation of the three phenotypes, cystic hyperplasia, solid acinar, and solid nodular malignancy, by the ADC-[¹⁸F]FDG combination (Fig. 4B and Supplementary Table S2).

The identification of the solid nodular phenotype was an important finding and an example of the benefits of combined PET/MRI, as this nodular malignancy was not identifiable with single-modality imaging. In the MRI it was distinguished from the background and identified as a solid region, but could not be differentiated from the acinar region, whereas [¹⁸F]FDG could not detect the lesion at all. Thus, if a comparable masked phenotype is present in human disease, the differentiation of this phenotype from metabolic active solid phenotypes could only be provided by the combined imaging procedure of PET/MRI.

To confirm the translational potential of the described approach, we investigated five patients with biopsy-proven IBC-NST, using simultaneous PET/MRI 7 ± 3 days before surgery. After surgical excision, pathologic analysis of the whole lesions was performed.

The first approach was the metabolic identification of tumor tissue within the breasts of the patients, to verify the potential of the method to identify malignant breast tissue, as described for the preclinical model. For this purpose, VOIs covering both the entire healthy and the entire diseased breasts were drawn on the basis of the CE-MRI. The clustering approach for [¹⁸F]FDG was applied separately on the sum of all voxels within the healthy (Fig. 5A) and within the diseased breasts (Fig. 5B). The best fit for the healthy breasts was identified for a sum of four Gaussian distributions, whereas the diseased breasts were best fitted by a number of five Gaussian distributions. The highest uptake population within the diseased breasts, including SUV > 1.1, was completely absent in the healthy breasts (Fig. 5A and B).

Applying this threshold of SUV > 1.1 to the whole breasts, the tumors within the breasts of all five patients were identified and confirmed as malignant tumors by histologic analysis (Fig. 5E, first row). The dimensions of the tumors given by this approach (maximal diameter, exemplary tumor: 16 × 15 × 15 mm³; Fig. 5E) showed a good correlation of the identified tumors as compared with the dimensions described by postsurgical pathologic analyses (maximal diameter, exemplary tumor: 16 × 13 × 10 mm³; Fig. 5E and Supplementary Table S3). This first approach could provide a user independent method to noninvasively identify malignant tumors.

As the main purpose of the study was the intratumoral classification of the lesions, we isolated all voxels within the morphologically identified tumor regions, to use the intratumoral clustering on the sum of all tumor voxels within the five primary tumors. For [¹⁸F]FDG, the clustering approach identified five Gaussian distributions, separated by the following thresholds: SUV = 0.5, 0.9, 1.6, and 2.6 (Fig. 5C). The ADC data were clustered by only two Gaussian distributions as the best fit, separated by the threshold of 1.1 × 10⁻³ mm²/s, distinguishing a low diffusion population from a high diffusion population (Fig. 5D). However, in this case, the populations were not clearly separated, yielding a high amount of a mixed population within the blue range with low ADC values.

The thresholds were plotted back into the MRI data to illustrate the clustered populations spatially within the VOIs (Fig. 5E, third row). In accordance to the preclinical data, the clinical data showed the complementary pattern of [¹⁸F]FDG uptake and ADC. Although the [¹⁸F]FDG uptake was higher in solid tumor regions, the ADC values were low in the solid tumor regions. However, higher ADC values were observed in the healthy gland structure of the breast and in single tumor regions, which were therefore suggested to be less solid.

As the previous approach already showed that [¹⁸F]FDG uptake values up to SUV of 1.1 represent the background uptake of healthy breast tissue, the first two populations (SUV < 0.5 and SUV 0.5–0.9) could not be distinguished from the background of the healthy tissue. As these populations were only present in the periphery of the tumors, surrounding the actual tumor areas, detected by the CE-MRI, they could be neglected as background populations (Fig. 5E, third row).

The remaining three uptake populations, especially the two highest populations, were considered as real tumor populations and were spatially correlated to the histologic findings (Fig. 6).

Comparing different regions of aggressiveness within the tumors scored by their proliferative activity and invasive growth pattern by histology, highly aggressive IC regions with high mitotic activity showed the highest [¹⁸F]FDG accumulation represented by the highest uptake population, which is marked in red in the figures (SUV > 2.6).

Here, we present the tumors of three representative patients. Figure 6A shows a tumor consisting homogeneously of an invasive carcinoma, according to histology, which is mainly represented by the highest red [¹⁸F]FDG population in the PET data. However, less aggressive regions with lower mitotic activity showed lower [¹⁸F]FDG accumulation, represented by the green (SUV 1.6–2.6) or blue (SUV 0.9–1.6) uptake populations. The tumor in Fig. 6B contained large areas of sclerosis with less tumor cells and less mitotic activity that correlated with the green uptake population. In contrast, sporadic small invasive carcinoma regions of high proliferation and aggressiveness within this tumor were represented by small spots of the red [¹⁸F]FDG population in the imaging data (Fig. 6B). Focusing only on the CE-MRI scans, both tumors in Fig. 6A and B had very similar morphology. The inter- and intratumoral biological and histological differences of these two tumors could only be distinguished by the metabolic PET clustering.

Moreover, combining [¹⁸F]FDG-PET, CE-, and DW-MRI, several regions in a tumor with a very heterogeneous morphologic and histologic pattern, composed of multiple foci of invasive carcinoma and a fibroadenoma component, were distinguished (Fig. 7). Although the extent of the tumor tissue was delineated by the CE-MRI, the aggressive spots of the invasive carcinoma were identified as described above by the high [¹⁸F]FDG uptake population in red (SUV > 2.6), that also correlated with low ADC values (Fig. 7, Slice 1). In contrast, the fibroadenoma could be distinguished from the rest of the tumor by a lower [¹⁸F]FDG accumulation (blue population, SUV 0.9–1.6) combined with the high ADC population (ADC > 1.1 × 10⁻³ mm²/s; Fig. 7, Slice 2).

This clinical data demonstrated that a comprehensive tumor profiling was possible using combined PET/MRI in the clinical setting. Thus, the method showed great potential for the clinical translation in larger scale studies.

Current trends in patient care aim for individualized medicine, making diagnostics and treatment strategies more complex. An important prerequisite for individualized therapies is a sophisticated diagnostic strategy that reveals even small alterations in the disease profile.

The standard diagnostic procedure in breast cancer involves mammography, sonography, and histologic validation. However, there are limitations to these techniques, for example in young patients with dense breast tissue (26). The value of PET and MRI for breast cancer diagnosis remains controversial. As recommended by the national comprehensive cancer network (NCCN) in high-risk patients, MRI has clear advantages (27–30), but an improvement of patient outcome was not yet confirmed (28, 31). Although not considered in the NCCN guideline to date, there is growing evidence for the value of PET and positron emission mammography (PEM; refs. 32–36). Avril and colleagues correlated [¹⁸F]FDG uptake with several prognostic markers such as histologic tumor type (ductal vs. lobular), microscopic tumor pattern (nodular vs. diffuse), and tumor proliferation (37). Yet, standard [¹⁸F]FDG-PET appears insufficient as single information for breast cancer diagnosis (38). Other promising tracers, e.g., addressing hormone-receptor status or gastrin-releasing peptide receptor are currently under clinical investigation (32, 39–41). Major restrictions in most clinical studies compared with this study are the exclusive use of single modality imaging and conventional image analysis. For example, when mammography is difficult, as in high-risk patients with mammographic dense breast tissue (42), the excellent soft tissue contrast of the MRI and the high sensitivity of the PET and functional MRI can provide essential information in the diagnostic process (26, 37, 43). Here, we showed that in combination with advanced analysis approaches, molecular and functional imaging can provide a more detailed characterization of the tumors inter- and intratumorally.

The assessment of tumor heterogeneity represents an emerging challenge in cancer diagnostics and treatment, for example, polygenic drug resistance is a major issue in cancer patient care to date (44). Potts and colleagues summarized outstanding work identifying breast cancer heterogeneity and predicting therapy response and resistance on the basis of immunohistochemistry (4). To overcome the snapshot character of these biopsy-based diagnoses, in vivo diagnostic tools like MRI, CT, and PET, combined with advanced analysis approaches, such as multispectral analysis or texture analysis, which are described as a powerful prognostic feature to improve diagnostics, tumor staging as well as therapy response assessment in many types of cancer (45–47) must be correlated to histology. Especially MRI, including functional and spectroscopic techniques showed great potential in imaging intratumoral heterogeneity and identifying biological subpopulations within single tumors, distinguishing for example viable tumor from necrosis and adipose healthy tissues (48), or classifying the tumor microenvironment in terms of vascular heterogeneity, hypoxia, and acidity (49, 50).

In our preclinical study, only combined PET/MRI and advanced image analysis enabled the detection and discrimination between two malignant tumor phenotypes (solid nodular and acinar) from premalignant cystic hyperplasia. Especially the identified solid nodular tumor regions with very low [¹⁸F]FDG uptake would be misinterpreted as nonmalignant using a single PET approach, if detected at all.

In the translational study with five IBC-NST tumor patients, the simultaneous PET/MRI approach including the GMM analysis led to the differentiation of highly aggressive and mitotic active tumor regions to less aggressive and sclerotic regions as well as to benign lesions such as a fibroadenoma. These results pointed out the opportunities of in vivo classification of tumors in the diagnostic workflow.

Using this analysis, phenotypic identification is based on data-derived thresholds, yielding a biology-driven distinction rather than user-dependent classification, showing prospective potential, where stable thresholds derived from the training data set were successfully transferred to other tumors of the same origin in the preclinical setting. Once defined, identified, and verified, the thresholds may be applied to other tumors within the same disease without the need of the complete image analysis procedure for every individual dataset. Its applicability and diagnostic value in other preclinical models as well as a larger scale validation in the clinical setting, however, was beyond the scope of this work. This analysis describes that the coexistence of different phenotypes within the same tumor can be visualized by multiparametric [¹⁸F]FDG PET/MR imaging, in both in animals as well as in patients suffering from breast cancer. This new approach has to be tested in further prospective studies; we distinguished a clinical significance to our approach.

Because treatment stratification is based on tumor biology received by minimal invasive biopsy, an accurate tissue sampling, especially in multifocal and multicentric breast tumors, is highly desirable. Therefore, our noninvasive tumor classification might enable a precise target selection and biopsy guidance in breast cancer, by revealing the presence of tumor heterogeneity and detecting the most aggressive part of the tumor, and thus would reduce the risk of undergrading and undertreatment.

Another application of this multiparametric imaging might be the early response evaluation after one to two cycles of neo-adjuvant chemotherapy (NACT), in order to adapt cancer treatment regimen or to spare ineffective and toxicity-related therapy. Because a complete remission after NACT is related with longer survival, an accurate prediction of a pathologic complete response after NACT and prior to surgery is highly desirable and might have a significant impact on further treatment stratification, for example mastectomy versus breast conserving therapy especially in multifocal/multicentric breast cancer. Finally, in cases of larger resection volumes, our approach might facilitate the final histopathological assessment and enhance the diagnostic accuracy, by defining point of interest with suspicious vital/avital tumor areas that would be processed separately after the surgical procedure.

Similar combined [¹⁸F]FDG/ADC approaches showed high diagnostic value, for example, in lung and breast cancer patients (51, 52) but could not provide prognostic thresholds due to the exclusive use of single value analysis in the breast cancer study and high grades of heterogeneity in the investigated lung disease, although analyzed by a GMM approach.

Current trends in cancer research showed that radiomics produce promising prognostic markers when applied on large patient cohorts (16, 17, 53). Our work further demonstrates that combined PET/MRI may complement these approaches by adding another biology-derived feature into the workflow, potentially moving one step further toward personalized medicine and individualized therapy planning.

No potential conflicts of interest were disclosed.

Conception and design: J. Schmitz, J. Schwenck, Q. Chen, B. Wietek, A. Staebler, R.D. Cardiff, B.J. Pichler, A.M. Schmid

Development of methodology: J. Schmitz, J. Schwab, Q. Chen, M. Hahn, B. Wietek, O.H. Aina, A.D. Borowsky, K. Nikolaou, B.J. Pichler, A.M. Schmid

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J. Schmitz, J. Schwenck, M. Hahn, B. Wietek, N. Schwenzer, A. Staebler, N.E. Hubbard, A.D. Borowsky, S. Brucker, K. Nikolaou, R.D. Cardiff, B.J. Pichler

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): J. Schmitz, J. Schwab, J. Schwenck, M. Hahn, A. Staebler, N.E. Hubbard, R.D. Cardiff, B.J. Pichler, A.M. Schmid

Writing, review, and/or revision of the manuscript: J. Schmitz, J. Schwab, J. Schwenck, L. Quintanilla-Martinez, M. Hahn, B. Wietek, N. Schwenzer, A. Staebler, N.E. Hubbard, G. Reischl, A.D. Borowsky, S. Brucker, K. Nikolaou, C. la Fougère, B.J. Pichler, A.M. Schmid

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): J. Schwab, Q. Chen, G. Reischl, A.D. Borowsky, R.D. Cardiff, A.M. Schmid

Study supervision: B. Wietek, S. Brucker, R.D. Cardiff, A.M. Schmid

Other (performed histological analysis and interpretation of mouse model): L. Quintanilla-Martinez

Other (analysis, interpretation and taking pictures of the mouse histopathology): U. Kohlhofer

Other (provided the PyVmT MINO model, provided training in use of transplants, and analyzed the histopathology. Visited Tuebingen to discuss the results and to help with their analysis. Dr. Aina was a postdoctoral trainee who assisted and was paid from my grants.): R.D. Cardiff

The authors would like to thank Maren Harant, Funda Cay, Daniel Bukala, Sandro Aidone, Carsten Groeper, Gerd Zeger, Holger Schmidt, Kerstin Fuchs, Laura Kübler, and Judith E. Walls for excellent technical assistance and the radiopharmacy team for the extensive tracer syntheses. Further, the authors thank Siemens Healthcare, Erlangen, Germany, for providing the dedicated four-channel breast MRI coil for the clinical measurements.

Parts of this work were funded by fortüne grant 2223-0-0, the Swiss Werner Siemens Foundation, the European Research Council grant 323196 (ImageLink), and by grants U01 CA141582 and U01 CA141541 from the National Cancer Institute's Mouse Models of Human Cancers Consortium.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.