The established role of hypoxia-induced signaling in prostate cancer growth, metastasis, and response to treatment suggests that a method to image hypoxia in tumors could aid treatment decisions. Here, we present consumption and supply-based hypoxia (CSH) imaging, an approach that integrates images related to oxygen consumption and supply into a single image. This integration algorithm was developed in patients with prostate cancer receiving hypoxia marker pimonidazole prior to prostatectomy. We exploited the intravoxel incoherent motion (IVIM) signal in diagnostic diffusion-weighted (DW) magnetic resonance (MR) images to generate separate images of the apparent diffusion coefficient (ADC) and fractional blood volume (fBV). ADC and fBV correlated with cell density (CD) and blood vessel density (BVD) in histology and whole-mount sections from 35 patients, thus linking ADC to oxygen consumption and fBV to oxygen supply. Pixel-wise plots of ADC versus fBV were utilized to predict the hypoxia status of each pixel in a tumor and to visualize the predicted value in a single image. The hypoxic fraction (HFDWI) of CSH images correlated strongly (R2 = 0.66; n = 41) with pimonidazole immunoscore (HSPimo); this relationship was validated in a second pimonidazole cohort (R2 = 0.54; n = 54). We observed good agreement between CSH images and pimonidazole staining in whole-mount sections. HFDWI correlated with tumor stage and lymph node status, consistent with findings for HSPimo. Moreover, CSH imaging could be applied on histologic CD and BVD images, demonstrating transferability to a histopathology assay. Thus, CSH represents a robust approach for hypoxia imaging in prostate cancer that could easily be translated into clinical practice.
Significance: These findings present a novel imaging strategy that indirectly measures tumor hypoxia and has potential application in a wide variety of solid tumors and other imaging modalities.
Graphical Abstract: http://cancerres.aacrjournals.org/content/canres/78/16/4774/F1.large.jpg. Cancer Res; 78(16); 4774–85. ©2018 AACR.
Tumor hypoxia is associated with metastatic disease and treatment resistance of prostate cancer (1–4). A method to assess hypoxia at diagnosis can help to select patients for intensified treatment and to avoid overtreatment of indolent, low-risk disease (5). This would meet an urgent need for improvement of the current treatment decision system. Oxygen electrodes and biopsy-derived techniques have been used to assess hypoxia in previous work (1–4), but these methods are hampered by sampling errors due to multifocality and intratumor heterogeneity. Medical imaging is a highly appealing approach, because hypoxia can be visualized throughout the entire prostate and repeated measurements can be performed.
No imaging method for hypoxia in prostate cancer has so far reached clinical use. Positron emission tomography (PET) with hypoxia-specific tracers has been successful in several cancer types (6), but has failed to detect hypoxia in this disease (7). Magnetic resonance imaging (MRI) approaches like blood oxygenation level–dependent (BOLD) imaging or related techniques to measure oxygenation directly are challenging and far from routine use (8–10). Moreover, indirect hypoxia measures, including vascular parameters derived from dynamic contrast enhanced (DCE)-MRI (11), have not shown the specificity needed for a hypoxia biomarker in prostate cancer, as demonstrated in preclinical work (12). Hence, in addition to an impaired oxygen supply detected in such images, hypoxia is promoted by a high oxygen consumption in tumor regions with increased cellularity (13). A better approach to visualize hypoxia may therefore be to combine information on both oxygen consumption and supply into a single image.
Diffusion-weighted (DW)-MRI could be an appropriate method for assessing both oxygen consumption and supply in prostate tumors. Based on images with different diffusion weighting (b-values), the diffusion of water molecules can be quantified by calculating the apparent diffusion coefficient (ADC). ADC images have been shown to reflect cellularity in prostatic tissue and are used in the diagnostic evaluation to distinguish malignant from nonmalignant lesions (14). In addition to diffusion, water molecules move through the tissue by fluid flow in the blood vessels. By using the intravoxel incoherent motion (IVIM) model, the DW-MR signal from the intravascular water molecules, being transported by blood flow, can be separated from the signal arising from the extravascular diffusing water molecules (15). This commonly neglected part of the DW signal may be used to create images of the fractional blood volume (fBV; refs. 16–18). Integration of ADC and fBV images to assess hypoxia would be a novel approach that has not been addressed in any cancer type.
Here, we aimed to develop a method to visualize hypoxia in prostate cancer through a pixel-wise integration of images reflecting oxygen consumption and supply into a single hypoxia image. This consumption and supply-based hypoxia (CSH) imaging technique was developed by exploiting the IVIM model in analysis of diagnostic, multiparametric DW images of patients who received the hypoxia marker pimonidazole before radical prostatectomy. Matched comparison of DW images and histology in whole-mount sections from the prostatectomy specimens was performed to investigate whether ADC and fBV correlated with cell density (CD) and blood vessel density (BVD), respectively, and thereby were related to oxygen consumption and supply. Matched comparison was also performed to calibrate the integration algorithm for hypoxia imaging against pimonidazole staining. We further demonstrated the robustness of CSH imaging by validation in an independent cohort and by showing its transferability to a histopathology assay for CD and BVD. The potential of CSH imaging in the treatment decision of prostate cancer is illustrated by showing strong associations between the hypoxic phenotype visualized in the images and postsurgical clinicopathologic markers.
Patients and Methods
In total, 114 patients referred to the Norwegian Radium Hospital for radical prostatectomy and enrolled into our ongoing FuncProst study (NCT01464216) were included (Supplementary Table S1). Most patients had intermediate or high-risk disease according to the D’Amico risk classification (19). All patients underwent preoperative multiparametric MRI, in accordance with the European Society of Urogenital Radiology (ESUR) 2012 Guidelines (20). In addition, a comprehensive DWI sequence with 11 different diffusion gradient strengths (b-values) was included. Patients received the hypoxia marker pimonidazole prior to surgery. Pimonidazole hydrochloride (Hypoxyprobe Inc.) was administered intravenously (i.v.) in the first patients enrolled. Due to a production stop of the intravenous administrable pimonidazole by the manufacturer, pimonidazole was given orally to the last included patients. The patients were divided into two cohorts, cohorts 1 and 2, depending on whether they were enrolled during the period of i.v. or oral pimonidazole, respectively (Supplementary Table S1).
A three-armed robotic DaVinci system (Intuitive Surgical) was used to perform radical prostatectomy (21). Pelvic lymph node dissection was performed in most high-risk patients and when preoperative MRI raised suspicion of lymph node metastasis. Radiologically, lymph nodes were evaluated based on node diameter and morphology in isotropic 1 mm T2-weighted (T2W) images. Lymph node status was determined either by pathologic examination of the nodes or considered negative if preoperative MRI was negative and serum level of prostate-specific antigen (PSA) at 6 weeks after prostatectomy was undetectable.
Grossing of the prostatectomy specimens was performed according to a standardized protocol (22). A strict experimental setup was used to ensure that histology and MR image analysis was performed in the same part of the prostate (Fig. 1). First, an experienced radiologist (KHH) predicted index tumor lesion based on multiparametric MRI. After surgery, by guidance of palpation and the MRI report, the prostate was cut horizontally into two halves where the assumed index tumor was located. Two punch biopsies were taken from the tumor for other purposes than the subject of this study. The prostate specimen was fixed in 10% buffered formalin for at least 48 hours and embedded in paraffin blocks. Based on 5-μm-thick hematoxylin and eosin (H&E)-stained whole-mount sections of each block, tumor was outlined by an experienced uropathologist (AKL, LV), and histopathologic staging and grading were performed according to the TNM classification (23) and an upgraded Gleason score system (24). Tumor size was quantified as the largest extent of the index lesion based on two perpendicular diameters. In case of multifocal tumor, index lesion was defined as, in descending order, pathologic T stage, Gleason score, and tumor size (25). Sections of the block with the largest index tumor area were used for defining tumor in MR images and for histologic analysis. The study was conducted in accordance with the Declaration of Helsinki and approved by the institutional review board and the Regional Committee for Medical and Health Research Ethics in South-East of Norway (2010/1656). Written informed consent was achieved from all patients.
MRI was performed on a 1.5 T General Electric Discovery 450 magnet with a 32 channel phased array coil (GE Medical Systems). The imaging protocol consisted of morphologic, T1- and T2W sequences covering the pelvis and the lower abdomen, and functional, DW (b = 1500) and DCE sequences covering the pelvic region. These sequences were solely used to outline and visualize the tumor. In addition, a comprehensive DW image sequence was recorded, using a transversal 2D spin echo echo-planar imaging (SE-EPI) sequence covering the entire prostate with a field of view (FOV) 180 mm × 180 mm, repetition time (TR) of 3,000 ms, echo time (TE) of 59.4 ms, echo train length of 92, bandwidth of 1,953 Hz/pixel, number of averages (NEX) of 2, 3 diffusion directions, slice thickness of 4 mm, slice gap of 1 mm, and in plane imaging matrix resolution of 0.70 × 0.70 mm2. Diffusion weighting was performed with b-values of 0, 100, 200, 300, 400, 500, 600, 700, 800, 900, and 1,000 s/mm2. To minimize bias in the DWI parameters caused by possible gradient nonlinearity (26), the patients were placed as close to the magnet isocenter as possible. Total acquisition time for the DW sequence was 4 minutes and 9 seconds.
DW image analysis
The DW images were analyzed with a simplified version of the IVIM model (15). The IVIM model assumes that the DWI signal originates from two compartments and can be described by a biexponential equation (Fig. 1):
where the first term represents signal loss due to intravascular movement of water, characterized by the pseudodiffusion parameter D*, and the second term represents signal loss due to extravascular diffusion of water, characterized by the diffusion coefficient ADC. The fractional blood volume, fBV, represents the fractional volume of the intravascular space, and S0 and S(b) are the image signal intensity obtained without (b = 0) and with diffusion weighting, respectively. Equation A can be simplified by assuming that the signal loss with increasing b-value caused by intravascular pseudodiffusion is fast compared with the signal loss due to extravascular diffusion (D*>>ADC). For large b-values, the signal loss therefore mainly depends on the extravascular diffusion, and Eq. A is reduced to a monoexponential equation, termed the simplified IVIM model:
The assumption, D*>>ADC, is fulfilled for b-values where the signal loss is linear in a logarithmic plot:
Hence, ADC and fBV can be calculated by using an adequate number of high b-values to accurately determine ADC together with a b-value of zero to obtain S0.
ADC and fBV images were produced with in-house made software developed in Matlab (MathWorks). Briefly, ADC and fBV were calculated for each pixel by fitting Eq. C to plots of ln(S(b)/S0) versus b, using a linear least square fit algorithm (Fig. 1). The b-values from 200 to 800 s/mm2 were chosen based on linearity analysis of ln(S(b)/S0) (Supplementary Method S1). The goodness of fit was assessed with Pearson correlation coefficient, R, and pixels with R2 < 0.9 were excluded. In 3 patients, poor fitting occurred for more than 20% of the tumor pixels, and these patients were excluded. In addition, 5 patients were excluded due to body movement within the diffusion series or because a different DWI sequence was applied. Consequently, 50 patients from cohort 1 and 56 patients from cohort 2 were included in the DW image analysis. This procedure led to ADC and fBV values that were comparable with the data achieved with the IVIM model, not influenced by the accuracy of the model fit and not confounded by nonbiological cross-correlation between the parameters (Supplementary Method S1).
Pimonidazole hydrochloride (Hypoxyprobe Inc.) was administered 13 to 24 hours before prostatectomy at a dose of 500 mg per m2 body surface. There was no relationship between pimonidazole staining in histologic sections and time from pimonidazole administration to complete dissection of the prostate (Supplementary Fig. S1). In cohort 1, pimonidazole was applied in 100 mL 0.9% NaCl i.v. over 20 minutes to 43 patients, whereas 60 patients in cohort 2 received pimonidazole orally in tablets. Totally 9 patients in cohort 1 and 2 patients in cohort 2 received no pimonidazole due to either anaphylactoid reaction to MR contrast agent or absence of study personnel, or correct quantification of pimonidazole staining was not possible due to postponing of surgery.
Pimonidazole immunohistochemistry was performed on the neighboring section of the H&E section used to define tumor for the MR image analysis, using a monoclonal mouse antibody for pimonidazole (clone 220.127.116.11, 1:100, Hypoxyprobe Inc.). The sections were stained using the DAKO EnVision TM Flex + System (Dako). The PT-Link (Dako) and EnVision TM target retrieval solution at high pH were used for antigen retrieval. The sections were incubated with the antibody for 30 minutes. Pimonidazole staining was visualized by 3,3-diaminobenzidine (DAB), and the sections were counterstained with hematoxylin. Pimonidazole antibody concentration was titrated to achieve high dynamic in staining intensity across the tumors. A tumor found to be positive was selected as a positive control and included in each run. As a negative control, the antibody was substituted with mouse myeloma protein of the same concentration and subclass as the pimonidazole antibody.
The staining pattern of the malignant glands was evaluated by an experienced pathologist (AKL, LV) blinded to MRI and other immunohistochemistry data, as described (4). Fractions of nuclear and moderate to strong cytoplasmic staining were determined separately and given immunoscores from 0 to 5 (0: 0%; 1: 1%–10%; 2: 11%–50%; 3: 51%–90%; 4: 91%–100%; and 5: 100%). The average value of cytoplasmic and nuclear immunoscore was used as the pimonidazole-based hypoxia score (HSPimo). Interobserver reproducibility for the quantification of immunoscore was evaluated previously on frozen biopsy samples and found to be good (κ = 0.80; ref. 4).
Histology images of CD and BVD
In cohort 1, 41 patients with paired pimonidazole and DWI data were chosen for histologic analysis of CD and BVD. A neighboring section of the pimonidazole section was stained for endothelial cells using a CD31 polyclonal rabbit antibody (ab28364, 1:50, Abcam) and the same protocol as for pimonidazole staining. CD31 staining was visualized by DAB, and the sections were counterstained with hematoxylin to visualize cell nuclei. Six whole-mount sections were not of sufficient quality for automated digital analysis and were excluded, yielding 35 patients for quantification of CD and BVD.
The stained sections were imaged on AxioImager Z1 ApoTome microscope system (Carl Zeiss) equipped with a 5×/NA0.13 lens and a 1ccc1 ccd camera (Carl Zeiss), yielding a pixel size of 1.4 μm. The entire tumor was imaged with 10% overlap between images using an automatic stage equipped to the microscope. The images were stitched together using AxioVision 4.8.2 (Carl Zeiss). To ensure reliable quantitative measures, all sections were imaged in one session using the same microscope acquisition settings.
Quantification of CD and BVD was performed automatically by an in-house developed program written in Matlab, as described (Supplementary Method S2). Briefly, blood vessels and cell nuclei in the tumor region were segmented based on DAB and hematoxylin staining, respectively. The cell nuclei and endothelial area fractions were thereafter presented in CD and BVD images. Because pixel-wise comparison between DW and histology images was not appropriate due to the difference in thickness between the images (4 mm vs. 5 μm) and the large heterogeneity of prostate cancer at the microregional level, median CD and BVD of the tumor was used in analysis against DWI data and HSPimo.
Curves were fitted to data by regression analysis. The Pearson correlation test was used to search for significant linear correlations between parameters. Statistical comparisons of data were carried out with Student t test when the data complied with the conditions of normality and equal variance. Under other conditions, comparisons were carried out with Mann–Whitney rank-sum test. When comparing more than two groups, the one-way ANOVA test was applied. For comparison of two nominal variables the χ2 test was used. Probability values of P < 0.05 were considered significant. The statistical analyses were conducted using SPSS (IBM Corp).
Pimonidazole staining is associated with postsurgical clinicopathologic markers
Pimonidazole staining intensity and distribution varied considerably both within the tumors and among patients (Fig. 2A and B). Focal pimonidazole staining was also seen in benign glands in some patients (Fig. 2B), as reported previously by Hoskin and colleagues (27). A median HSPimo of 2 and 1.5 was measured in cohorts 1 and 2, respectively. HSPimo was significantly higher in the former cohort (P < 0.005; Fig. 2A), possibly because this cohort contained more tumors with aggressive features (Supplementary Table S1). HSPimo was significantly associated with several clinicopathologic markers, including postsurgical obtained parameters like pathologic tumor grade (P < 0.00005), pathologic Gleason score (P < 0.05), lymph node status (P < 0.005), and largest tumor extent (P < 0.00005; Fig. 2C; Supplementary Table S2), suggesting that a high HSPimo reflects an aggressive phenotype of prostate cancer. Pimonidazole staining therefore seemed to be an appropriate hypoxia measure for the development of the CSH imaging technique. No association was found between HSPimo and systemic blood level of PSA at diagnosis (Supplementary Table S2).
DWI parameters reflect CD and BVD in whole-mount sections
To investigate whether the DWI parameters ADC and fBV correlated with CD and BVD, respectively, we compared the ADC and fBV images with the corresponding CD and BVD images in 35 cohort 1 patients with paired DWI and histology data. The histology images were constructed with a pixel size equal to that obtained in the DW images to facilitate visual comparison (Supplementary Methods S2), as demonstrated for 2 patients in Fig. 3A, although a large difference in thickness between the DW images and histologic sections may contribute to differences at the microregional level. Intratumor heterogeneity was seen in all images of both patients; however, patient 1 had generally high ADC and low CD compared with patient 2 (Fig. 3A). In line with this example, a strong negative correlation was found between ADC and CD in analysis of all patients (R2 = 0.46, P < 0.0001; Fig. 3B). Patient 1 also showed higher fBV and higher BVD than patient 2 (Fig. 3C), and a strong positive correlation was found between these two parameters (R2 = 0.44, P < 0.0001; Fig. 3D). Both ADC and fBV showed a significant, but weaker, relationship to HSPimo (R2 = 0.18 and R2 = 0.35, respectively; Supplementary Fig. S2). There was no significant correlation between ADC and fBV (P > 0.05). These results support our hypothesis that ADC and fBV can be used as measures of CD and BVD in prostate tumors. Moreover, the assumption that the DWI parameters provide different pathologic information related to oxygen consumption and supply, respectively, appears to be reasonable.
CSH images provide measures of hypoxic fraction and visualize hypoxia
Based on the above assumption, an algorithm for combining images of oxygen consumption and supply into a single hypoxia image was constructed from the ADC and fBV images of cohort 1 patients (Supplementary Document S1). In short, pixel-wise plots of fBV versus ADC were generated for each tumor. Most pixels from hypoxic tumors, according to HSPimo, were presented down to the left corner in such plots compared with pixels from less hypoxic tumors, as exemplified with a more hypoxic and less hypoxic tumor in Fig. 4A. Hence, from the relationship between DWI and histology parameters (Fig. 3), these pixels most likely represent tumor regions with low oxygen supply and high consumption, and therefore regions with a high probability of being hypoxic. In contrast, pixels closer to the upper right corner represent regions with high oxygen supply and low consumption, and with a low probability of being hypoxic. A curve discriminating pixels from hypoxic and non-hypoxic regions could be approximated to a linear curve expressed as
where ADC0 and fBV0 are the intersections of the line with the horizontal and vertical axes, respectively (Fig. 4A; Supplementary Document S1).
Fraction of pixels below the discrimination line was used as a measure of the hypoxic fraction (HFDWI) predicted by CSH imaging. The optimal line was defined in an iterative procedure involving all 41 cohort 1 patients with paired DWI and pimonidazole data to achieve the best possible prediction of HSPimo, as assessed by the correlation between HFDWI and HSPimo. The strongest correlation was found for a line with ADC0 = 0.79 × 10−3 mm2/s and fBV0 = 0.43 (Supplementary Document S1). In comparison, the median ADC and fBV for the patients were 0.71 × 10−3 mm2/s and 0.12, respectively. Based on this line, the correlation between HSPimo and HFDWI (R2 = 0.66, P < 0.000001; Fig. 4B) was much stronger than the correlation between HFPimo and ADC (R2 = 0.18) or fBV (R2 = 0.35; Supplementary Fig. S2). Moreover, by using an HSPimo cutoff of 2 for classifying a tumor as hypoxic based on the median HSPimo in cohort 1, the area under curve (AUC) in the receiver operating curve (ROC) analysis was as high as 0.95 (Fig. 4C).
To visualize the hypoxia information in an image, the distance from a pixel to the optimized discrimination line was used to calculate a hypoxia score, HSDWI, of each pixel (Supplementary Document S1). The score, thus, represents the probability of the pixel to be hypoxic and was color coded to reflect the dynamics in the score (Fig. 4D). There was a good visual agreement between the CSH images and pimonidazole staining in whole-mount sections, as demonstrated in Fig. 4E and F for the 2 patients shown in Fig. 3.
CSH imaging is validated in an independent cohort and relates to clinicopathologic markers
The robustness of the CSH technology was evaluated by analyzing fBV-ADC plots in 54 cohort 2 patients with paired DWI and pimonidazole data. The discrimination line optimized in cohort 1 with an ADC0 of 0.79 × 10−3 mm2/s and an fBV0 of 0.43 was used to calculate HFDWI of each tumor. A significant correlation was found between HSPimo and HFDWI (R2 = 0.54, P < 0.000001; Fig. 5A) that was much stronger than the correlation between HFPimo and ADC (R2 = 0.13) or fBV (R2 = 0.27; Supplementary Fig. S2). Moreover, an AUC of 0.88 was achieved in ROC analysis with the same HFPimo cutoff of 2 that was used in cohort 1 (Fig. 5B). The CSH images showed good concordance with pimonidazole staining in cohort 2, as demonstrated by the visualization of two clearly hypoxic regions by both imaging and pimonidazole staining in Fig. 5C and D. These results confirmed the validity and, thus, suggested robustness of our method, both for providing a measure of the hypoxic fraction and for visualizing hypoxia. The two cohorts showed an almost equal linear relationship between HSPimo and HFDWI (Supplementary Fig. S3), showing that the relationship was not dependent on whether pimonidazole was administered i.v. or orally.
To investigate whether CSH imaging provided hypoxia measures with the same relationship to clinicopathologic markers as HSPimo (Fig. 2C; Supplementary Table S2), cohorts 1 and 2 were merged. Patients with lymph node metastasis had significantly higher HFDWI than patients without lymph node metastasis (Fig. 6), in line with the pimonidazole data. HFDWI was also associated with all other postsurgical clinicopathologic markers that were significant for HSPimo, including pathologic tumor stage (P < 0.00001), pathologic Gleason score (P < 0.0005) and largest tumor extent (P < 0.0005). Moreover, the associations to clinicopathologic markers were stronger for HFDWI than for ADC or fBV (Supplementary Table S3).
The CSH imaging method shows transferability to a histopathology assay
To explore whether our method could be transferred across assays, we used the integration algorithm directly on histology-derived CD and BVD images, assuming that these images reflected oxygen consumption and supply, respectively (Supplementary Document S1). Pixel-wise plots of 1-CD versus BVD were generated, as demonstrated in Fig. 7A for the same 2 patients as presented in Figs. 3 and 4. The linear curve discriminating hypoxic and non-hypoxic pixels was expressed as
where BVD0 and (1-CD)0 are the intersections of the line with the vertical and horizontal axis, respectively. The same iterative procedure described for the DWI data was used to optimize the discrimination line to achieve the best possible prediction of HSPimo. The hypoxic fraction derived in the analysis, HFHist, showed a significant correlation to HSPimo (R2 = 0.59, P < 0.00001; Fig. 7B) that was stronger than the correlation between HSPimo and CD (R2 = 0.23) or BVD (R2 = 0.40; Supplementary Fig. S4). Moreover, CSH images created from the CD and BVD images as described for the DW images (Supplementary Document S1) showed a good agreement with pimonidazole staining in whole-mount sections (Fig. 7C–F). These results further validated our method and demonstrated its robustness across assays.
By pixel-wise integration of multiparametric DW images reflecting oxygen consumption and supply in prostate cancer, we generated hypoxia images that could identify aggressive disease. The importance of both consumption and supply for the hypoxic fraction in tumors has been well established (13); however, the principle of combining these factors has not been exploited in imaging before. Pimonidazole staining in whole-mount sections from prostatectomy specimens was utilized for the construction and validation of the CSH imaging technique, ensuring that the images reflected a true and clinically relevant hypoxia phenotype. Hence, the staining pattern was similar to previous observations in prostate tumors, regardless of whether pimonidazole was i.v. or orally administered (28). It is likely that pimonidazole, as applied here, detects mainly chronic hypoxia and not acute hypoxia (29). However, the staining was found to correlate with several aggressive features in this and previous work (4), and was comparable with oxygen electrode measurements associated with radiotherapy outcome (2), supporting its clinical relevance. DWI is widely used in the detection and staging of prostate cancer, and a diagnostic protocol with multiple b-values is recommended by ESUR (20). Our method would therefore be easy to implement in the clinic. CSH imaging could also be applied directly on histologic images of CD and BVD, suggesting application of the method on other assays as well for assessment of hypoxia in prostate tumors.
CSH imaging is based on the hypothesis that the hypoxic fraction can be derived from surrogate measures of oxygen consumption and supply. In addition to these parameters, the hypoxic fraction depends on the tolerance of tumor cells to oxygen-deprived conditions (30). A high tolerance promotes cell survival and reduces necrosis formation under hypoxia, leading to a higher hypoxic fraction than when the tolerance is low. The tolerance seems to be high in prostate cancer, as compared with many other cancer types (30). Hence, areas of necrosis are small in extent and seldom seen, according to our own observations and other reports (31), and large regions with oxygen concentrations below 0.1% can be found (30). The variability in hypoxic fraction across the tumors in our work was therefore probably not caused by differences in hypoxia tolerance. This conclusion is also supported by the successful integration of oxygen consumption and supply parameters into a robust measure of hypoxic fraction. However, for application of our method on cancer types developing necrosis, a possible variability in hypoxia tolerance among tumors could influence the results and should be considered.
For construction of the image integration algorithm, we assumed that CD and BVD could be used as surrogate measures of oxygen consumption and supply, respectively. These assumptions are based on a simplified description of the oxygen conditions in tumors that may not be valid for all cancer types. In addition to differences in CD, the oxygen consumption may differ among tumors due to differences in the respiration rate per cell, which is affected by the growth fraction and cellular proliferation rate (32). Mesko and colleagues (33) found that low ADC, and hence high cellularity, was associated with a high growth fraction in prostate tumors, suggesting that the CD and growth fraction covary in this disease. In addition to BVD, blood vessel functionality and oxygen content of the blood may vary within and between tumors and contribute to differences in oxygen supply. The suitability of BVD as surrogate marker of oxygen supply has been thoroughly studied in preclinical models, summarized by Lee and colleagues (34), but has not been investigated in prostate cancer patients. Our results showed a correlation of both BVD and fBV with hypoxic fraction, suggesting that low BVD is a limiting factor for the oxygen supply in prostate tumors. In particular, each of the CD and BVD parameters correlated with hypoxic fraction, and the combination of the two parameters led to a strong and robust measure of this fraction. It is therefore likely that CD and BVD are appropriate measures of oxygen consumption and supply, respectively, in prostate tumors. However, for transferability of the method to other cancers, other surrogate markers may be preferred.
Strong linear correlations were found between ADC and CD and between fBV and BVD, justifying the use of these DWI parameters for the development of CSH imaging. A relationship between ADC and CD has been reported in several studies on prostate cancer (35–37). The connection between fBV and histologic measures of vascularity, on the other hand, has scarcely been studied in a clinical setting. No results have been reported on prostate cancer; however, our findings are consistent with reports on pancreatic cancer and rectal cancer (17, 18). The strong correlation of the DWI parameters with histology implies that they can be determined with high accuracy. This led to a robust discrimination line and thereby measures of the DWI-based hypoxic fraction in our work, as confirmed by validation of the line in an independent cohort. To assess the intravascular part of the DWI signal, imaging at small b-values including a b-value of zero is required (38). IVIM imaging is not included in the ESUR guidelines, and the contribution of the intravascular signal to the image intensity has often been ignored in DWI protocols. Our study demonstrates a significant potential of the IVIM information together with conventional ADC to assess hypoxia in prostate cancer that motivates for implementation of the required b-values in the protocol.
CSH imaging may be useful for identifying and targeting hypoxia and thereby aid implementation of customized treatment strategies for patients with prostate cancer. The current treatment decision system is not able to identify all patients with aggressive disease in need of intensified treatment, and the CSH images may add to the current system. They may further facilitate accomplishment of safe and reliable clinical trials to counteract the adverse effect of hypoxia and improve the local control rate in radiotherapy of prostate cancer. By visualizing hypoxic regions within the tumor, the images can serve as basis for radiation dose escalation to these resistant subvolumes (39). In addition, a better understanding of the effect of hypoxia-targeted drugs in combination therapies can be achieved through the separate information provided by the ADC and fBV images. Hence, several promising radiosensitizers in prostate cancer, such as androgen deprivation or antidiabetic agents, inhibitors of the hypoxia-inducible factor HIF1 and drugs for vascular normalization, are likely to affect hypoxia by changing the tumor CD and/or BVD (40, 41).
In conclusion, CSH imaging provided hypoxia images and measures of the hypoxic fraction that may be of value in the treatment decision for prostate cancer patients. By showing its feasibility on DW images, implementation of CSH imaging into clinical practice is facilitated. The demonstration of a more general validity on histology images encourages similar investigations on other imaging modalities, including DCE-MRI, as well. Moreover, the possibility that our method could be used in other cancer types merits further investigation.
Disclosure of Potential Conflicts of Interest
S. Patzke is a research manager at Nordic Nanovector ASA. No potential conflicts of interest were disclosed by the other authors.
Conception and design: T. Hompland, K.H. Hole, B. Brennhovd, T. Seierstad, H. Lyng
Development of methodology: T. Hompland, K.H. Hole, B. Brennhovd, T. Seierstad, H. Lyng
Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): T. Hompland, K.H. Hole, H.B. Ragnum, E.-K. Aarnes, L. Vlatkovic, A.K. Lie, S. Patzke, B. Brennhovd, T. Seierstad, H. Lyng
Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): T. Hompland, K.H. Hole, B. Brennhovd, T. Seierstad, H. Lyng
Writing, review, and/or revision of the manuscript: T. Hompland, K.H. Hole, H.B. Ragnum, E.-K. Aarnes, L. Vlatkovic, A.K. Lie, S. Patzke, B. Brennhovd, T. Seierstad, H. Lyng
Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): T. Hompland, K.H. Hole, H.B. Ragnum, E.-K. Aarnes, L. Vlatkovic, A.K. Lie, S. Patzke, B. Brennhovd, T. Seierstad, H. Lyng
Study supervision: H. Lyng
Technical assistance from D. Trinh and A.A. Nyboen is highly appreciated. U. Ryg is acknowledged for valuable help with the graphical abstract. T. Hompland was supported by the South-Eastern Norway Regional Health Authority (grant no. 2015020). H.B. Ragnum was supported by the European Union 7th Framework program (grant no. 222741-METOXIA).
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.