Introduction

Breast cancer (BC) is the most common cancer and the second leading cause of death by cancer among women. This cancer has various molecular subtypes that differ in aggressiveness, treatment, and outcome [40, 53]. As a great heterogeneity regarding genetic profile, aggressiveness, and therapeutic options [44] persist, it appears essential to develop additional diagnostic tools to better characterize tumors and to guide therapeutic strategies, using biology, next-generation sequencing, and imaging. Tumors are often associated with an increase in glucose metabolism [32] and in the context of BC, baseline metabolism is positively associated with tumor aggressiveness [15]. Beyond glucose metabolism, tumors have an increased blood flow (BF), due to the angiogenesis [36]. Since a mismatch between tumor metabolism and tumor blood flow has been reported [31], evaluation of these two different parameters appears essential to fully characterize breast tumors. Different imaging modalities, such as dynamic contrast-enhanced magnetic resonance imaging (MRI), 15O-water positron emission tomography/computed tomography (PET/CT) and 1-h dynamic 18F-fluorodeoxyglucose (18F-FDG) PET/CT, have yielded kinetic parameters reflecting vascular permeability and perfusion. Methods based on the 2-min half-life 15O remain the gold standard for perfusion evaluation [26]. However, because an on-site cyclotron is needed, only a restricted number of research centers are able to use 15O-water. Furthermore, 18F-FDG PET has become a standard for primary staging of large and locally advanced breast cancer [9] and a full 1-h dynamic 18F-FDG PET acquisition allows tumor blood flow and metabolism evaluation after a single injection of 18F-FDG. As a 1-h acquisition is hardly suitable in clinical routine, an alternative method allowing the measurement of the tumor blood flow from a first-pass 18F-FDG PET has been developed by Mullani et al. [34]. The calculated blood flow [10, 20, 22] showed an excellent correlation with the tumor perfusion measured with 15O-water [35]. In the context of BC, few studies have suggested the interest of the perfusion evaluation to monitor BC therapies and its ability to provide predictive and prognostic information [10, 13, 20, 22, 30, 47].

Recently, investigation of tumor heterogeneity using texture analysis (TA) has emerged [45, 46]. Standard PET/CT parameters combined with TA [2, 3, 19] appear promising to bring valuable information on the relationship between metabolism heterogeneity and clinical, biological parameters, tumor subtypes, or tumor aggressiveness [16, 29, 33, 45, 46]. In a similar way, the present paper hypothesizes that perfusion heterogeneity analysis might bring valuable additional information to metabolic imaging with 18F-FDG PET/CT. Indeed, the reasons why tumors are heterogeneous are still largely unknown and one of the explanations could be the vascular disorders related to the anarchic nature of the tumor angiogenesis [1]. Nonetheless, although several previous studies have quantified 18F-FDG uptake intratumoral heterogeneity, to the best of our knowledge, none was carried out on tumor BF heterogeneity. The first-pass 18F-FDG PET/CT represents an opportunity to assess tumor perfusion in addition to tumor metabolism and to evaluate in a single exam, perfusion heterogeneity and its relation to metabolism heterogeneity.

The objective of this prospective study is to analyze the global tumor blood flow and its heterogeneity in a large cohort of 217 newly diagnosed breast cancer cases according to tumor biological characteristics and molecular subtypes. These perfusion parameters were compared to those classically derived from metabolic studies to further investigate links between perfusion and metabolism.

Materials and methods

Patients and study design

From January 2011 to May 2017, 217 patients with newly diagnosed stage II or III BC were prospectively recruited, based on oncologists’ recommendations for a 18F-FDG PET/CT scan before any treatment. Patients with high glycemia (> 9 mmol/l) and unable or unwilling to undergo PET scan were excluded. The institutional review board approved this prospective study as a current-care study. The medical team documented patient non-opposition in source documents and in the information notice provided to the patient, and the need for a signature indicating informed consent was not required. Core biopsies of primary tumors were used to evaluate histological characteristics, such as histological type and tumor grade, using the Scarff–Bloom–Richardson (SBR) system [6]. Tumor, lymph nodes, metastasis, and classification of malignant tumors (TNM) were established in accordance with the 8th edition of the American Joint Committee on Cancer (AJCC) staging system [4]. Molecular markers examined included estrogen receptor (ER), progesterone receptor (PR) and HER2 expression. ER and PR status were considered positive if tumors showed at least 10% of positive cells. HER2 status was assessed by immunohistochemistry or fluorescent in situ hybridization (FISH) and graded from 0 to 3 + . Scores of 3 + were considered positive. HER2 amplification confirmed by FISH was used to assess positivity in ambiguous cases (2 + ). Patients were classified into three groups according to their molecular subtypes: luminal (hormone receptors (HR)-positive, HER2-negative), HER2 (HER2-positive) and triple negative (TN) (HR-negative, HER2-negative).

PET-CT protocol

All images were acquired with a Gemini TruFlight PET/CT scanner (Philips Medical Systems, Eindhoven, The Netherlands) providing an axial field of view of 18 cm and a transaxial slice thickness of 4 mm. An automatic PET infusion system (Bayer Medical Care, Inc., Indianola, PA, USA) was used to inject a bolus of 3 MBq/kg of 18F-FDG. Simultaneously with injection, an early first 8-min 18F-FDG PET scan was performed, in a prone position and centered on the breast. A low-dose CT scan (120 kVp, automatic tube current modulation, 0.5-s rotation time, 16×1.5 mm collimation, a pitch of 0.7) of the same region was carried out for anatomical registration and attenuation correction. The dynamic first-pass was reconstructed based on twelve 10-s frames extracted from the first 2 min of the PET acquisition. A whole-body emission and transmission scans were acquired 60 min later, in a supine position. Finally, 90 min after injection, the delayed static image was acquired in a prone position with a two-step PET/CT scan restricted to the chest. The emission acquisition time was 4 min per bed position. All the PET images were reconstructed using a first 3D-ordered subset iterative (OSEM) time of flight reconstruction technique with three iterations and 33 subsets. Images matrix sizes of 144×144 with an isotropic voxel size of 4 mm were used. The emission data were corrected for random coincidences, decay, dead time, scattering and attenuation.

Parametric images

Measurement of blood flow

The tumor blood flow measurement using dynamic PET of first-pass 18F-FDG uptake is based on the one-compartment first-pass model of Mullani et al [34, 35] and has already been explained in detail in the literature [10, 34, 35].

Several volumes of interest (VOIs) were manually delineated on the early static PET image and the corresponding CT image using a dedicated workstation (Extended Brilliance Workspace 3.5; Philips): a first VOI representing the contours of the primary tumor, a second VOI into the ascending aorta, and a third VOI in the contralateral normal breast. The latter was used as a background reference. All the VOIs were superimposed on the dynamic PET image and the time–activity curve (TAC) of the ascending aorta was then calculated. Counting rate reached a maximum when the arterial input to the tumor ceased and when the tracer had not left the region of interest (both arterial input and venous drainage were zero). This time Tm corresponds to the end of the first pass of the tracer and was visually determined on the arterial TAC. A parametric BF image (Fig. 1), containing a BF value in each tumor voxel, was then computed with in-house scripts using the vv tools [43], which are based on the Insight Segmentation and Registration Toolkit (ITK) library [52], by using the following equation:

$$ BF\ =\frac { Q({ T}_{m} )} { E{\int}_{0}^{{ T}_{m} } { Ca(t)dt} } $$
(1)

where Q(Tm) is the amount of tracer measured in each tumor voxel at time Tm and Ca(t) is the arterial concentration of the tracer measured at time t. The extraction fraction E of 18F-FDG was assumed to be equal to one [21].

Fig. 1
figure 1

Parametric blood flow (ml/min/g) image (a) from 18F-FDG PET first-pass dynamic (10-s frames) acquisition and SUV image (b) from delayed (4-min per bed position) acquisition, for a 63-year-old woman with TN invasive ductal carcinoma of right breast. A.1: tumor VOI from parametric blood flow image after a rescaling using 32 grey levels. B.1: tumor VOI from SUV image after a rescaling using 64 grey levels. Primary tumor BF_max: 0.5, BF_mean: 0.2, SUV_max: 13.5 and SUV_mean: 8.1

Full size image

Measurement of glucose metabolism

Semi-automatic segmentation using a contrast-dependent method of metabolic active tumor volumes (MATVs) were performed on the delayed static PET images. This segmentation is based on the determination of an optimal threshold value as a function of the measured contrast in the image. It requires a calibration procedure described in detail by [42]. For this purpose, phantom acquisitions were performed using the NEMA/IEC body phantom with sphere-to-background activity concentration ratios ranging from 40:1 to 2:1. These ratios were chosen to span the whole range of contrast ratios observed in the delayed PET images. A series of six VOIs were drawn with a distance of at least 5 mm from the spheres. The mean standardized uptake values (SUV) of all these VOIs was used as the background value. All the spheres were automatically segmented by varying a threshold in steps of 0.01. From the optimal threshold values yielding the best matching with the true spheres’ volume (± 2%), a regression function was calculated to determine the relation with the normalized thresholds and the contrast ratios, as defined in [42]. Two regression functions were used, a first one for the small volumes (less than 2 ml) and a second one for the larger volumes (more than 2 ml). The adaptive threshold algorithm was computed by using in-house tools based on the vv software. All the necrotic tumor regions were included in the tumors VOIs. Late PET images were then converted to SUV (Fig. 1) by using the classical expression:

$$ SUV=\frac { Tumor\ activity\ concentration\ (MBq/ml)} { Injected\ dose\ (MBq)\ /\ body\ weight\ (g)} $$
(2)

PET-derived features

Standard PET parameters reflecting tumor metabolism and perfusion were extracted from the SUV PET and blood flow parametric images, respectively. The computed features were the average, (SUV_mean and BF_mean for the metabolism and the perfusion, respectively), the standard deviation, (SUV_sd and BF_sd) and the maximum (SUV_max and BF_max) tumor values. In addition, the metabolic active tumor volume and the total lesion glycolysis (TLG = MATV× SUV_mean) were calculated for the metabolism. All these first-order statistics provide global information on tumor perfusion and metabolism, without any spatial information on the voxel arrangement. On the contrary, TA using second-order statistics can locally describe and quantify the spatial arrangement of the voxels. Several studies have been conducted in order to determine robust and reproducible texture features (TF) from tumor metabolic PET images [18, 28, 50] but, to the best of our knowledge, none have been conducted on the tumor perfusion images from first-pass 18F-FDG PET in BC. Therefore, the same TFs were selected to study both metabolism and perfusion in order to make comparisons easier. Four TFs were used for the analysis of tumor heterogeneity: entropy, inverse different moment (IDM), correlation, and dissimilarity. The entropy is related to the orderliness in the image (how regular the pixel value differences are). This feature quantifies the randomness of the voxel intensity distribution.

$$ Entropy\ =\ -\sum\limits_{i=1}^{Nbins} { \sum\limits_{j=1}^{Nbins} { P(i,j)\ln { (P(i,j))} } } $$
(3)

The IDM, also called local homogeneity, quantifies the homogeneity within an image and will be high if the intensities of each pair of voxels are similar.

$$ Inverse\ different\ moment\ =\sum\limits_{i=1}^{Nbins} { \sum\limits_{j=1}^{Nbins} { \frac { P(i,j)} { 1+{ (i-j)}^{2} } } } $$
(4)

The correlation measures the linear dependency of grey levels and the dissimilarity quantifies the variation of grey-level voxel pairs.

$$ Correlation\ =\sum\limits_{i=1}^{{ Nbins} } { \sum\limits_{j=1}^{Nbins} { \frac { P(i,j)(i-{ \mu}_{i} )(j-{ \mu}_{j} )} { { \sigma}_{i} { \sigma}_{j} } } } $$
(5)
$$ Dissimilarity\ =\sum\limits_{i=1}^{{ Nbins} } { \sum\limits_{j=1}^{Nbins} { P(i,j)\left| i-j \right|} } \\ $$
(6)

These features were computed in 3D using grey-level co-occurrence matrices (GLCM), which take into account pairs of voxels arrangement. For each of the 13 spatial directions, a GLCM matrix was computed between consecutive voxels (i.e., a 1-distance relationship). P(i,j) is the probability value recorded for a pair of voxels in the co-occurrence matrix, μi and μj correspond to the averages on row i and column j, σi and σj are the variances on row i and column j and Nbins is the number of grey levels used in the image quantization. The TF values were calculated for each matrix and then averaged to obtain the final TF values.

All the texture parameters were calculated with a fixed number of 64 bins for the metabolism analysis as recommended in the literature [17, 18, 48, 50], and with a number of 32 bins for the perfusion analysis, to minimize the impact of the noise which is higher in parametric blood flow than in metabolic images. Original PET images were rescaled between the minimum and the maximum intensity within the tumor VOI (a relative rescaling) (Fig. 1).

Statistical analysis

Statistical analysis was performed with R [41]. For all the features extracted from the parametric images, average, median, and standard deviation values were calculated. Shapiros’ test was used to test the normality of the variable distribution. Since most distributions were not normal, relations between features and phenotypes were assessed using the Kruskal–Wallis test, considering a significance level of 0.05. If statistical significance was found, pairwise comparisons using a Mann–Whitney test with a Bonferroni adjustment was computed to find out differences between phenotypes. Associations between PET features and clinical and physiological parameters were investigated using a two-sided Mann–Whitney test. Correlations between PET features were evaluated with the Spearman rank test.

Results

Patient characteristics

The study was conducted on 217 patients with a median age of 49 years and an average primary tumor diameter of 34.5 mm. The median MATV was 6.1 ml with an interquartile range of 8.2 ml and the median perfused volume was 4.0 ml with an interquartile range of 7.1 ml. Ductal carcinoma represented 93.6% of the patients, 5.5% were lobular, and 0.9% were not classified. Seventy-nine patients were HER2-positive, 81 luminal/HER2-negative, and 57 had TN tumors. Patient and tumor characteristics are summarized in Table 1.

Table 1 Patient and tumor characteristics
Full size table

Correlation between PET features

Figure 2 shows the Spearman correlation heat map of PET features (metabolism and perfusion). Regarding perfusion features, BF_correlation, BF_entropy and BF_IDM were strongly correlated with each other (r> 0.74), whereas BF_dissimilarity was inversely correlated with the other perfusion textural features (r<-0.71). Except for a weak correlation between BF_max and BF_entropy (r= 0.61), BF_max and BF_mean were not correlated with perfusion textural features. Concerning metabolic features, M_correlation, M_IDM and M_entropy were strongly correlated with each other (r> 0.70) and with MATV (r> 0.73) and TLG, whereas M_dissimilarity was strongly inversely correlated with the other metabolic textural features, MATV, and TLG. SUV_max and SUV_mean were not correlated with the other metabolic features. Regarding the correlation between perfusion and metabolic features, except for a weak correlation between BF_max and TLG (r = 0.64), there was no significant correlation between standard metabolism features (SUV_max, SUV_mean, MATV, TLG) and perfusion standard features (BF_max and BF_mean). In contrast, BF_correlation, BF_entropy and BF_IDM were correlated with M_correlation, M_entropy, MATV, and TLG, and inversely correlated with M_dissimilarity. Finally, BF_dissimilarity was correlated with M_dissimilarity and inversely correlated with the other metabolic TFs.

Fig. 2
figure 2

Spearman rank correlation heat map between blood flow texture features, metabolism texture features, and standard PET features. The dark blue colors represent strong positive correlations (close to 1), the dark red represent strong inverse correlations (close to -1). The light/white represent an absence of correlation (close to 0). BF: blood flow; IDM: inverse different moment; M: metabolism; MATV: metabolic active tumor volume; SUV: standardized uptake value; TLG: total lesion glycolysis

Full size image

Relationship between PET-derived features and tumor characteristics

Table 2 summarizes the PET-derived features in association with the tumor characteristics. Concerning perfusion parameters, there was no difference according to ER, PR, and HER2 status. In contrast, BF_max, BF_entropy and BF_correlation were significantly higher in the event of T stage 3-4, lymph node involvement or ductal carcinoma, whereas BF_dissimilarity was lower. T stage 3–4 tumors also exhibited higher BF_IDM. Finally, BF_max was higher for SBR grade III. Regarding metabolic features, both SUV_max and SUV_mean were significantly higher in the event of SBR grade III, ductal carcinoma, ER negative, and PR negative. In contrast, no differences were found for SUV_max and SUV_mean concerning HER2 status, T-stage and N-stage. Higher MATV and TLG were found in patients with T-stage 3–4, N+ status, and with HER2- status. TLG was also higher in the event of SBR grade III and ductal carcinoma. M_entropy, M_lDM and M_correlation were higher in patients with T-stage 3–4 and HER2 negative, whereas M_dissimilarity was lower. M_entropy and M_correlation were also higher in patients with N+ status.

Table 2 Relationship between PET-derived features, histological factors, and tumor characteristics (Mann–Whitney p values). The average of the feature values for each group is presented in parentheses when significant
Full size table

Comparison between phenotypes

Table 3, Figs. 3, and 4 summarize the comparison of blood flow and metabolism characteristics between BC phenotypes. Regarding global BF analysis, no significant differences among phenotypes were observed. TF analysis showed no significant differences between HER2 and Luminal tumors. However, TN tumors had a significantly higher entropy than luminal tumors (entropy: TN = 6.18 ± 1.65, luminal = 5.49 ± 1.45, p= 0.011), and a significantly higher homogeneity than the HER2 ones (IDM: HER2 = 0.14 ± 0.05, TN = 0.17 ± 0.06, p= 0.015; correlation: HER2 = 0.29 ± 0.22, TN = 0.37 ± 0.23, p= 0.034)). Concerning metabolism analysis, TN tumors showed higher values of SUV_max and SUV_mean than HER2 and luminal tumors. In addition, TF analysis showed that HER2 tumors had a significantly lower homogeneity and entropy than luminal tumors (IDM: HER2 = 0.06 ± 0.02, luminal = 0.07 ± 0.02, p< 0.001; entropy: HER2 = 6.22 ± 1.53, luminal = 6.80 ± 1.59, p= 0.034). HER2 tumors had also a lower homogeneity and entropy than TN ones (IDM: TN = 0.07 ± 0.03, entropy: TN = 6.86 ± 1.93, p= 0.023 and p= 0.047). No significant differences in heterogeneity were reported between TN and luminal tumors.

Table 3 Comparison between blood flow and metabolism characteristics for each breast cancer phenotype
Full size table
Fig. 3
figure 3

Distribution of PET features among phenotypes for blood flow. IDM: inverse different moment

Full size image
Fig. 4
figure 4

Distribution of PET features among phenotypes, for metabolism. IDM: inverse different moment; SUV: standardized uptake value

Full size image

Discussion

The present paper shows that adding blood flow evaluation, from the first pass of 18F-FDG, to the more conventional delayed 18F-FDG uptake, brings new insights into breast cancer biology. Global blood flow increases in the context of large tumors and axillary lymph node invasion. Texture analysis shows that perfusion heterogeneity parallels that of metabolism. These results suggest a complementary role of advanced textural features and standard PET imaging parameters for the in vivo biological characterization of BC lesions, supporting the hypothesis that advanced PET imaging analysis can provide additional information.

As previously reported, global tumor perfusion was found to be almost identical in the three BC subtypes [21, 47], while glucose uptake significantly increased from luminal to HER2 and TN tumors [9, 10, 15, 16, 21, 33]. The increased metabolism of some tumors, especially the TN subtype, has led to the mismatch concept between perfusion and tumor metabolism [10, 21]. Regarding the relationship between global perfusion parameters and tumor characteristics, no association between blood flow average (BF_mean) and hormonal receptors or SBR was observed, as previously reported [10, 21, 33]. Although no perfusion difference was highlighted in a previous paper from our institution [20], a significantly higher perfusion was however noted, in the present study, for ductal carcinomas. The larger number of patients here (217 vs. 150) may potentially explain this difference in the results. As there is an important numeric disproportion between the lobular and the ductal tumor population, results must be carefully interpreted. However, the observation of a higher perfusion for ductal carcinomas is in line with some previous studies where the expression of vascular endothelial growth factor (VEGF) protein was significantly higher in invasive ductal carcinomas than in invasive lobular carcinomas [11, 27].

Moreover, in the present study, T3-T4 tumors were found to have a higher BF_max than T1-T2 tumors and this observation has not been reported before. It could be argued that tumor blood flow increases with size but the blood flow “max” is calculated within a small part of a tumor, i.e., a cube of 64 mm3 representing less than 100 mg of tumor. Yet, this observation is in line with a recent study [24] where a significantly lower microvessel density was associated with a smaller tumor size. Fifty percent of tumors with a diameter between 2 and 5 cm had a microvessel density > 72.9/mm2 by comparison to only 38% in tumors with a diameter smaller than 2 cm. Moreover, we noted that, in patients with lymph nodes involvement (N+), primary tumors also exhibited a higher BF. Lymph node involvement reflects tumor local invasion and is an important prognostic factor. Similarly, a borderline correlation has been reported [5] between vascular proliferation index and positive nodal status in a series of 110 patients with locally advanced breast cancer. Supporting these observations, a color Doppler ultrasound study showed in a large cohort of breast carcinomas (368 patients) that the peak systolic flow velocity was significantly associated with tumor size and the presence of axillary lymph node metastases [7]. In the same way, a direct association has also been reported between expression of VEGF and the tumor size and the number of metastatic lymph nodes [11].

The differences observed in perfusion depending on tumor size and lymph node involvement might reflect an increased angiogenesis. Indeed, it has been suggested that angiogenesis in breast cancer is regulated by vascular factors secreted by tumor cells and has a close relationship with axillary nodule metastasis [8]. Angiogenesis plays an important role in both tumor growth and metastasis of many types of cancers and is recognized as an important hallmark of oncogenesis [14]. This might provide an explanation to the higher perfusion observed in primary tumors with clinical lymph node involvement, as extensive angiogenesis could facilitate cancer metastasis. These differences in perfusion according to T-stage and N-stage are not paralleled by similar differences in tumor metabolism, as no significant differences were noted for both SUV_max or SUV_mean. This observation might suggest that studying perfusion might provide further insights in breast cancer invasiveness.

In recent years, several studies have investigated the heterogeneity of 18F-FDG uptake in different tumor types. Some research has focused on breast cancer and various image-derived features, reflecting heterogeneity, have been proposed. Most of these studies have been mainly descriptive, attempting to find correlations with tumor subtypes and treatment responses. In the present study, using a single 18F-FDG PET/CT exam, we were able to characterize both tumor perfusion and metabolism heterogeneity and look for possible links between these two tumor hallmarks’ heterogeneity. Texture feature interpretation can be influenced by the method used to calculate it. Depending on the PET image rescaling method [28, 38], it was shown that the texture feature will not have the same behavior, and potentially not the same meaning [28, 29, 37,38,39]. Bearing this question in mind, a limited number of texture features known to be robust and reproducible [17, 19, 29, 38, 39, 50] were selected in the present study. Moreover, similar texture features were used to study both perfusion and metabolism, thus making comparisons easier.

To the best of our knowledge, no prior study has addressed perfusion heterogeneity using PET/CT in BC. Among a large population covering different BC phenotypes, the differences found between HER2 and TN tumors were significant, with a higher heterogeneity in the HER2 subtype. This difference was also found between HER2 tumors and luminal tumors, but was not significant. Metabolism heterogeneity analysis highlighted similar differences between phenotypes, with HER2 tumors significantly more heterogeneous than luminal and TN tumors. Consequently, HER2 subtype appeared here, as the most heterogeneous phenotype, in both perfusion and metabolism. It is worth noting that perfusion heterogeneity matched metabolism heterogeneity.

One possible interpretation of these facts might be related to the heterogeneous distribution of vascular density within the tumor, due to angiogenesis [1]. Indeed, previous published works underlined the relationship between the HER2-positive tumors and angiogenesis [23, 25, 51]. Moreover, for this BC subtype, clinical trials have suggested that addition of an antiangiogenic treatment to a conventional neoadjuvant chemotherapy, can increase the proportion of patients who achieve a pathological complete response [12], supporting the link between HER2 and angiogenesis.

Finally, unlike global perfusion parameters that show only a limited correlation with standard metabolism features [10, 47], perfusion TFs and metabolism TFs were highly correlated (for example, r= 0.84 for the correlation between M_entropy and BF_entropy). As the heterogeneity is assessed by comparison of neighbor pixels, this could reflect some anarchic distribution of perfusion and metabolism. One study has investigated the heterogeneity of microvessels in BC [1] at the microscopic level. The authors concluded that heterogeneity is even more important inside a given tumor than among various types of tumors. Results obtained here may suggest that understanding perfusion heterogeneity at the pixel level could possibly be a step forward in the comprehension of metabolism heterogeneity.

Study limitations

One of the main limitations of this study is the small volume of the tumors. However, small tumor size is one of breast cancer’s characteristics and the choice of TFs was done taking into account this aspect, using local TFs only. Considering the wide range of techniques available and the contradictory results, a careful interpretation of heterogeneity should be done. Regarding perfusion and metabolism tumor heterogeneities comparison, we performed a global rather than a voxel-to-voxel comparison given that perfusion and metabolism were evaluated from two different acquisitions.

No correction for partial volume effect was applied and the low count statistics in the dynamic images might add a bias that could affect the results [49]. By improving the signal-to-noise ratio and the spatial resolution of the dynamic short time frames, better image quality should be obtained, leading to new or more significant results. Improvement of spatial resolution and sensitivity provided by new digital PET or dedicated breast PET scanners should improve the analysis in this range of tumor sizes. Moscoso et al. [33] recently investigated the dedicated breast PET resolution improvement. They showed that this technique enables a more accurate heterogeneity characterization, but they focused only on metabolism analysis.

Conclusions

In this study, a significantly higher perfusion in tumors with lymph node involvement has been highlighted, whereas no significant difference was noted in tumor metabolism. In addition, tumor blood flow heterogeneity was different in different breast cancer phenotypes and this was not previously observed with a global analysis of the perfusion. A combination of standard PET parameters and textural features of both tumor perfusion and metabolism could represent a step forward towards a better understanding of tumor behavior and therapeutic strategies adaptation, in the era of personalized medicine. The next step would be to investigate if all these new texture features and their early changes during therapy might improve clinical outcome prediction or treatment response, as potential biomarkers of therapeutic response.