MRI Distinguishes Tumor Hypoxia Levels of Different Prognostic and Biological Significance in Cervical Cancer

Solid tumors show a highly heterogeneous oxygen distribution with hypoxia levels ranging from mild to moderate and severe (1). The hypoxia level determines resistance to cancer therapies like radiation, chemotherapy, and many molecular targeting drugs (1–4), and may therefore have large therapeutical consequences. Current understanding of how the different levels drive cancer progression and affect treatment response is scarce and mostly based on experimental studies (5–7). At mild hypoxia, around 2% O₂, activation of the hypoxia-inducible transcription factor HIF1 promotes metabolic reprogramming and cell survival (8, 9). More severe levels, below 1% O₂, may impair cell proliferation and lead to genomic instability (10, 11), and below 0.5% O₂, the cytotoxic effect of radiation is more than 2-fold reduced (2). Hypoxia may also induce epithelial–mesenchymal transition (EMT) and immune evasion of tumor cells (12, 13), but the levels of importance for these processes are not known. In patient tumors, earlier investigations using invasive electrodes to measure oxygen partial pressure (pO₂) have shown considerable differences across cancer types in the level most strongly associated with treatment outcome, ranging from 2.5 to 10 mmHg or approximately 0.3%–1.3% O₂ (14). More recent clinical work has almost exclusively focused on the presence or absence of hypoxia (15), mainly because oxygen electrodes are not feasible and alternative approaches to assess hypoxia levels are lacking. A method based on medical imaging would facilitate investigations of how individual levels relate to treatment outcome and tumor biology in patients, and help development of more efficient therapies to combat hypoxia.

Hypoxia occurs in tumors due to impaired oxygen supply by a chaotic vascular network and/or elevated oxygen consumption in regions with high cellularity (1). We recently presented a tool for pixel-wise combination of images reflecting oxygen consumption with images reflecting oxygen supply into images representing hypoxia (16). The consumption and supply-based hypoxia (CSH)-imaging tool was originally developed in patients with prostate cancer, using images of the apparent diffusion coefficient and fractional blood volume derived from diffusion weighted (DW) MR images. The information in the two images was utilized to reflect the difference between oxygen consumption and supply and thereby the probability of each pixel to locate in a hypoxic region. Although only the presence of hypoxia was addressed in this study, it is likely that a difference between oxygen consumption and supply within a tumor region also would provide information on hypoxia level. The CSH-principle may therefore be a basis for establishing an imaging approach for hypoxia levels.

Locally advanced cervical cancer is a disease for which better biological understanding and therapeutical approaches to overcome hypoxia are urgent (17, 18). In this work, we aimed to construct images that reflect a continuous distribution of hypoxia levels in cervical tumors by applying the CSH-tool. Our approach was based on dynamic contrast enhanced (DCE)-MRI, because this modality is state-of-the-art diagnostics for the disease. We showed that the DCE-MRI parameters ν_e and K^trans from the Tofts pharmacokinetic model (19) reflected oxygen consumption and supply, respectively, and could be successfully combined to generate hypoxia images in xenograft and patient tumors. We further developed an algorithm to assign surrogate measures of hypoxia levels to all pixels. The algorithm was validated by comparison with hypoxia levels reflected in pimonidazole-stained sections in xenograft tumors and hypoxia-related gene expression in patient tumors. The power of this approach was demonstrated by presenting the distribution of hypoxia levels in tumors of 74 patients and distinguishing levels associated with treatment outcome, a set of cancer hallmarks and stabilization of HIF1A protein.

Totally 74 patients with locally advanced cervical carcinoma, prospectively recruited to our chemoradiotherapy protocol, were included (Supplementary Table S1). Gene expression profiles and a gene score reflecting hypoxia were available from previous work (20) for 63 patients, and paraffin-embedded tissue sections for IHC were available for 73 patients. The gene score was based on the expression level of six hypoxia-responsive genes and increased with increasing amount of hypoxia (20). All patients received external radiotherapy combined with cisplatin (40 mg/m² weekly), intracavitary brachytherapy and follow-up as described previously (20). The study was approved by the institutional review board and the Regional Committee for Medical Health Research Ethics in southern Norway. Written informed consent was obtained from all patients.

HeLa and SiHa cervical cancer cell lines from ATCC were used. Confirmation of cell line identity and cell culturing were performed as described previously (21). Mycoplasma testing was conducted regularly and always prior to cryopreservation. Totally 1.5 × 10⁶ HeLa and 1.7 × 10⁶ SiHa cells in passages 5–20 after thawing were reseeded in 10-cm plastic dishes 24 hours before exposure to hypoxia at 0.2%, 0.5%, 1%, 2%, and 5% O₂ for 24 hours at 37°C, all with 5% CO₂, by using an Invivo₂200 chamber (Ruskinn Technology Ltd.). Normoxic controls (95% air, 5% CO₂) were included for all hypoxia experiments.

HeLa and SiHa cervical cancer xenograft tumors were established in female nude mice, bred at the animal department of our institute and kept in specific pathogen-free environment, with food and water supplied ad libitum. Totally 1 × 10⁶ HeLa cells or 2 × 10⁶ SiHa cells in passages 5–8 after thawing were suspended in 20 μL or 40 μL of Hank balanced salt solution and injected intramuscularly in both hind legs of adult mice. Tumor growth was monitored with anatomic T₂-weighted MRI. At the day of DCE-MRI, the hypoxia marker pimonidazole (60 mg/kg; Hypoxyprobe, Inc.) was administered intraperitoneally prior to MR scanning in 16 HeLa and 12 SiHa tumors. After the scan, 90–120 minutes after pimonidazole injection, the mice were euthanized by cervical dislocation, and the tumors were excised, formalin-fixed, and paraffin-embedded for IHC. All procedures were approved by the Norwegian Animal Research Authority and performed in accordance with guidelines on animal welfare of the Federation of Laboratory Animal Science Associations.

DCE-MRI of xenograft tumors was performed at a volume of 100–800 mm³, using a 7.05 T Biospec bore magnet (Bruker) and a fast bolus injection of 5.0 mL/kg body weight Gd-DOTA (Dotarem, Guerbet; Supplementary Method S1). Totally 8 images prior to and 57 images postinjection of Gd-DOTA were acquired with an axial T₁-weighted spoiled gradient recalled sequence (SPGR). The images had a spatial resolution of 234 × 234 × 1,000 μm³. The three most central tumor slices were analyzed.

In patients, DCE-MRI was performed at diagnosis, using a 1.5 T Signa Horizon LX tomograph (GE Medical Systems) with a pelvic phased array coil and a fast bolus injection of 0.1 mmol/kg body weight Gd-DTPA (Magnevist, Schering; Supplementary Method S1). Totally 1–2 series prior to and 12–13 series postinjection of Gd-DTPA were acquired with an axial T₁-weighted SPGR sequence. The images had a pixel size of 780 × 780 μm², slice thickness of 5 mm, and slice gap of 1 mm. All slices containing tumor were analyzed.

The tumors were outlined in T₂-weigthed MR images and coregistered with DCE-MR images. Pharmacokinetic analysis of contrast uptake curves obtained from the DCE-MR images was performed on a pixel-by-pixel basis using Tofts model (19) (Supplementary Method S1), and parametric images of K^trans and ν_e were generated. To construct hypoxia images, the CSH-tool was applied on the K^trans and ν_e images as described for DW-MRI (16). Hence, a pixel-wise plot of K^trans versus ν_e was generated for each tumor, representing decreasing oxygen consumption on the horizontal ν_e-axis and increasing oxygen supply on the vertical K^trans-axis. To determine a threshold for hypoxia, a line discriminating pixels in hypoxic and nonhypoxic regions, and thus defining the hypoxic fraction, was determined in an iterative procedure with all tumors, using an independent hypoxia measure as learning variable. The hypoxic fraction was calculated for each tumor and iteration, and correlated with the independent hypoxia measure. The optimal line was determined by the highest Pearson correlation coefficient and was described by its intersections with the horizontal (ν_e0) and vertical axes (K₀^trans).

Adjacent sections, 4–5 μm thick, from xenograft tumors were stained for hypoxia (n = 28), using a pimonidazole polyclonal rabbit antibody (1:3,500; Hypoxyprobe Inc.) and endothelial cells (n = 26), using a CD31 rabbit polyclonal antibody (1:50, ab28364; Abcam). Hematoxylin was used as counterstain to visualize cell nuclei. Digital histopathology was performed to quantify hypoxic fraction (HF_Pimo), cell density (CD), and blood vessel density (BVD; Supplementary Method S1). Sections from 73 patient tumors were stained with the monoclonal mouse HIF1A antibody clone 54 (1:25, no. 610958; BD Transduction Laboratories), as described previously (21). Percentage of HIF1A-positive tumor cells was scored manually based on nuclear staining: 0, 0%; 1, 1%–10%; 2, 11%–25%; 3, 26%–50%; 4, 51%–75%; and 5, >75%.

Gene expression profiling of HeLa and SiHa cells exposed to hypoxia at 0.2%, 0.5%, 1%, 2%, and 5% O₂ and normoxia (95% air) was carried out using Illumina bead arrays HT-12 v4 (Illumina Inc.). Total RNA was isolated using miRNeasy MiniKit (Qiagen). Complementary RNA was synthesized, labeled, and hybridized to the arrays. Signal extraction and quantile normalization were performed using software provided by the manufacturer (Illumina Inc.). The data were deposited in the Gene Expression Omnibus (GEO; GSE147384). Normalized gene expression profiles of 63 patients, generated previously using Illumina bead arrays WG-6 v3 (Illumina Inc.; ref. 20), were downloaded from GEO (GSE72723).

To compare hypoxic fractions derived from MR images and pimonidazole-stained sections, an adapted version of Pearson correlation test for similarity between two datasets was applied (22):

where x and y are sets of hypoxic fractions from MRI and pimonidazole, respectively, var(x) and var(y) are set variances and n is the set size of 28 tumors. The function is equal to 1 when the hypoxic fractions from the two modalities are perfectly correlated with a slope of 1. In cases of poor correlation or a slope deviating from 1, the similarity decreases.

Curve fitting was performed by regression analysis. Student t test was used for comparison of parametric data between groups, otherwise, Wilcoxon rank-sum test was applied. Linear associations were estimated by Pearson correlation. Association between hypoxic fraction and gene expression data was based on Spearman rank correlation. A cut-off P-value of 0.05 was used to achieve an appropriate number of genes for further analysis. To avoid unreliable associations caused by outliers, the analysis was repeated after removing the lowest and highest expression value of each gene. Gene set enrichment analysis (GSEA) was performed on the basis of 50 hallmark gene sets in the Molecular Signature Database, using an adjusted P-value of q < 0.05 to evaluate statistical significance. Clinical endpoint was progression-free survival defined as time from diagnosis to disease-related death or first occurrence of relapse. Patients were censored at their last appointment or at 5 years. Cox univariate proportional hazard analysis was performed, and Kaplan–Meier curves were compared using log-rank test. P-values of P < 0.05 were considered significant unless otherwise stated. The statistical analyses were performed using SigmaPlot and SPSS.

The possibility to construct hypoxia images from DCE-MR images was investigated in xenograft tumors by first examining whether ν_e and K^trans could be used to reflect oxygen consumption and supply. Images of ν_e and K^trans displayed resemblance with those of the histopathology parameters CD and BVD, respectively, with some disagreement possibly due to a 200-fold difference in slice thickness (Supplementary Fig. S1A and S1B). Consistent with these observations, histogram analyses based on the DCE-MR and histopathology parametric images of all tumors showed negative correlations between ν_e and CD and positive correlations between K^trans and BVD for almost all percentiles, including the median values (R² = 0.17, P = 0.03 and R² = 0.46, P < 0.0005, respectively; Supplementary Fig. S2A and S2B). No significant correlation was found between ν_e and BVD or between K^trans and CD (Supplementary Fig. S3A). Hypoxic fraction determined by pimonidazole staining (HF_Pimo) was correlated with both ν_e (R² = 0.46, P < 0.00005) and K^trans (R² = 0.22, P < 0.05; Supplementary Fig. S3B). ν_e and K^trans therefore seemed to be connected to hypoxia and contain different information related to oxygen consumption and supply, respectively, in line with other DCE-MRI studies using low molecular weight contrast agents (23).

On the basis of the above results, we searched to construct hypoxia images in xenograft tumors by combining images of ν_e and K^trans and using HF_Pimo as independent measure of hypoxia. In pixel-wise plots of K^trans versus ν_e, pixels from tumors having a high HF_Pimo were in general located more toward the lower left corner than pixels from tumors with a low HF_Pimo (Fig. 1A and B), consistent with the CSH-principle. The line that best discriminated pixels in hypoxic and nonhypoxic regions for all tumors combined was determined (Supplementary Fig. S4A). Pixels below the optimal line were considered hypoxic and the fraction of these pixels, HF_MRI, was strongly correlated to HF_Pimo (R² = 0.57, P < 0.000005; Fig. 1C). This correlation was stronger than between ν_e or K^trans and HF_Pimo (Supplementary Fig. S3). The resulting binary hypoxia images showed strong resemblance to the pimonidazole-stained sections (Fig. 1B and D).

To confirm applicability of the CSH-tool to produce hypoxia images in patient tumors, the same procedure was applied to pixel-wise plots of K^trans versus ν_e generated from the clinical images (Supplementary Fig. S4B and S5A). Similar to what we observed in xenografts, pixels from hypoxic tumors appeared to be located toward the lower left corner in these plots (Supplementary Fig. S5B). By using the gene score from previous work (20) as independent hypoxia measure, an optimal line to discriminate pixels in hypoxic and nonhypoxic regions for all tumors combined was determined (Supplementary Fig. S4B), and a HF_MRI was calculated for each tumor. A strong correlation between HF_MRI and hypoxia gene score (R² = 0.27, P < 0.00001; Fig. 1E and F) was found. This correlation was stronger than between ν_e or K^trans and gene score (Supplementary Fig. S6A). In analysis of all 74 patients, HF_MRI was strongly correlated with progression-free survival, where patients with high HF_MRI had poor outcome compared with the others (P = 0.0014; Supplementary Fig. S6B), consistent with the prognostic significance of the gene score (20). The correlation to outcome was weaker for K^trans or ν_e (P = 0.015 and P = 0.074, respectively; Supplementary Fig. S6B). All together, this showed that hypoxia images could be constructed using the DCE-MRI parameters ν_e and K^trans as input to the CSH-tool.

On the basis of the hypoxia images, an algorithm was developed that transferred the ν_e and K^trans information of each pixel into a single value representing the hypoxia level. A dimensional reduction approach was utilized, where each point in the plot of K^trans versus ν_e was projected onto the axis orthogonal to the optimal discrimination line. The use of this approach was based on a hypothesis that the location of a pixel in the plot; that is, the distance from the pixel to the optimal discrimination line, depends on the hypoxia level of the corresponding tumor region (Fig. 2A). This hypothesis is likely because the line represents the weighted information of K^trans (oxygen supply) and ν_e (oxygen consumption) underlying the level of the independent hypoxia measure. The hypoxia level, HL_MRI, can thus be expressed as:

where the level of the optimal line, described by the intersection points ν_e0 and K₀^trans, was set to zero, and increasing values of HL_MRI indicated more severe hypoxia. Application of the algorithm to calculate four hypoxia levels is shown in Fig. 2B, together with the underlying HL_MRI image (Fig. 2C).

A procedure to extract hypoxia levels from pimonidazole-stained tumor sections in xenografts was developed for validation of the algorithm. In vitro studies have shown that the binding efficacy of pimonidazole during hypoxia increases exponentially with decreasing oxygen concentration (24). In line with this, the pimonidazole-staining intensity was generally strongest close to necrotic regions (anoxia) and decreased with increasing distance from necrosis (Fig. 2D), most likely reflecting a hypoxia gradient. We therefore assumed that the staining intensity was proportional to hypoxia level, and produced pimonidazole-based images reflecting hypoxia levels (HL_Pimo) for the validation (Fig. 2E and F; Supplementary Method S1). Visual inspection showed large resemblance between the HL_MRI and HL_Pimo images (Fig. 2C and F), although there was a considerable difference in slice thickness between the two modalities. By this inspection, we further found that the staining intensity in pimonidazole-based images could be evaluated down to a HL_Pimo of 0.38. Below this limit, the intensity was weak with small changes, probably reflecting nonhypoxic levels.

Hypoxia levels from MR- and pimonidazole-based images (Fig. 2C and F) were compared in 28 xenografts. By varying the threshold for HL_MRI, from 0.11 in severe hypoxia to −0.05 at the mildest level, and for HL_Pimo, from 2.6 at the strongest staining intensity to 0.01 in the weakly stained region, we generated sets of hypoxic fractions (% of tumor >HL_MRI or HL_Pimo) for both modalities and all xenografts (Fig. 3A). The two datasets, each consisting of 28 × 200 hypoxic fractions, were first compared using similarity analysis, where we for each HL_MRI threshold identified the HL_Pimo threshold that led to the highest similarity between hypoxic fraction derived by the two modalities (Fig. 3B). Overall, the similarity values were high (>0.6) and an exponential relationship was observed between similarity-matched HL_MRI and HL_Pimo, presented as a linear relationship in a log-plot in Fig. 3C. This relationship is consistent with the exponential binding of pimonidazole with decreasing oxygen concentration (24). Correlation analysis of the most similar hypoxic fractions indicated how well HL_MRI reflected different hypoxia levels. A strong correlation (P < 0.001) was found for HL_MRI in the range of −0.03 to 0.1 (Fig. 3C), showing that hypoxic fraction from a large range of levels could be measured. Within this range, the mean hypoxic fraction based on all 28 xenograft tumors displayed considerable differences, ranging from 0.38 at mild hypoxia (HL_MRI = −0.03; Fig. 3D) to 0.07 at more severe hypoxia (HL_MRI = 0.06; Fig. 3E) and 0.02 at the most severe levels (HL_MRI = 0.1). These results supported that our algorithm to image hypoxia levels was reliable. Furthermore, the MRI-defined hypoxia levels could distinguish a large range of hypoxic fractions in xenograft tumors.

Because pimonidazole-stained histologic sections were not available for the clinical cohort, we constructed an indicator of hypoxia levels based on the expression of hypoxia-responsive genes to confirm the validity of our algorithm in patient tumors. We utilized that genes may be activated and, thus, show increased expression, at specific oxygen concentrations (25). Nine indicator genes were selected among 31 previously identified hypoxia-responsive genes in cervical cancer (21; Supplementary Document S1). The genes are regulated by HIF1 (AK4, PFKFB4, P4HA2), by both HIF1 and the unfolded protein response (STC2, ERO1A) or by unknown mechanisms (UPK1A, KCTD11, SNTA1, PYGL). By exposure of SiHa and HeLa cells to oxygen concentrations in the range of 0.2%–21% O₂, the concentration for half-maximal response was recorded for each gene (Fig. 4A and B), in a similar way as described for stabilization of HIF1A protein (8). This cell line–derived hypoxia activation level ranged from 0.55% to 1.81% O₂ (Fig. 4C; Supplementary Document S1). The HIF1A targets AK4 and PFKFB4 had the highest level, consistent with an HIF1 activation level of 1.5%–2.0% O₂ (8). Thus, the indicator genes showed a range of levels likely to be found in human tumors (6) and broad enough for testing our algorithm.

HL_MRI images were constructed for all 74 patient tumors (Fig. 4D). Using the same strategy as for xenografts, a set of 200 hypoxic fractions was calculated for each tumor using HL_MRI thresholds ranging from 0.1 in severe hypoxia to −0.3 as the mildest level. Expression data of the nine indicator genes were further retrieved from the gene expression profiles of each tumor. A correlation analysis of the two datasets was performed, where we for each indicator gene identified the HL_MRI threshold that led to the strongest association between hypoxic fraction and expression (Fig. 4E; Supplementary Document S1). These HL_MRI thresholds showed a strong correlation to the cell line–derived hypoxia activation level of the nine indicator genes (Fig. 4F; R² = 0.84, P < 0.0005). Although oxygen concentrations for half-maximal response in cell lines are not directly transferable to patient tumors, this relationship together with the above xenograft results strongly supported that HL_MRI provided a continuous, linear surrogate measure of hypoxia levels in tumors.

The relationship presented in Fig. 4F provided a tool to relate MRI-defined hypoxia levels in individual patient tumors to levels derived in cell lines. Aided by the relationship, we defined approximate HL_MRI intervals for severe, moderate, and mild hypoxia to characterize the hypoxia level distributions (Fig. 5A). The definitions corresponded roughly to those proposed by others (6). Median HL_MRI of all tumors combined was −0.08. This value was related to a cell line–derived level of 1.3% O₂ (Fig. 5A) and within the moderate hypoxia range. Moreover, the median value differed considerably across tumors, ranging from −0.22 (2.3% O₂) in mild hypoxia to 0.004 (0.8% O₂) in moderate hypoxia. A pie chart of each tumor was generated to visualize these differences, showing fraction of pixels within HL_MRI intervals of 0.05 (Fig. 5B; Supplementary Fig. S7). Most tumors contained a range from severe to nonhypoxic levels, however, fraction of the different levels varied considerably across patients.

To address whether differences in the pie charts were associated with differences in chemoradiotherapy outcome, the dataset of 200 hypoxic fractions for HL_MRI thresholds in the range of −0.3–0.1 used in Fig. 4E and F were included in survival analysis. The strongest association to outcome was found for hypoxic fractions below a HL_MRI threshold of 0.01 (Fig. 5C), which was related to a cell line–derived level of 0.7% O₂ and in moderate hypoxia, close to the interval of severe hypoxia (Fig. 5A). Hence, patients with a high hypoxic fraction below this level had a poor outcome compared with the others (P = 0.0014; Fig. 5D). In contrast, weaker or no association to outcome was found for more severe hypoxia; that is, the highest HL_MRI values, or for milder hypoxia.

The dataset of 200 hypoxic fractions used above was further correlated with gene expression profiles of the patient tumors to identify possible associations between hypoxia level and biological processes. Totally 1,344 genes with the strongest positive correlation for one or more HL_MRI thresholds were subjected to GSEA, yielding 36 significantly enriched hallmarks (Supplementary Table S2) that included 350 of the genes. By assigning the HL_MRI threshold showing the strongest correlation between hypoxic fraction and expression for the 350 genes, a distribution of hypoxia levels was produced for each of the 36 enriched hallmarks. In general, the individual HL_MRI distributions covered a large range of hypoxia levels, and most hallmarks (n = 26) had a median HL_MRI in the moderate hypoxia range, including well-known hypoxia-related processes like hypoxia and glycolysis (Supplementary Fig. S8 and S9).

The HL_MRI distributions were further compared across the 36 hallmarks, to search for differences in the distributions across biological processes. All hallmarks were tested against each other, and those with a difference (P < 0.05) to less than 25% of the others were removed to simplify analysis. For the remaining 15 hallmarks, three groups with a significant difference in HL_MRI distribution were identified (Fig. 6A; Supplementary Fig. S8). A group with the IFNα and IFNγ response hallmarks was associated with mild hypoxia (Fig. 6A and B). A group including G₂–M checkpoint, MYC targets, oxidative phosphorylation and MTORC1 signaling was related to significantly more moderate levels, whereas hallmarks like TNFA signaling via NFκB, DNA repair, inflammatory response, angiogenesis, and EMT were associated with the most severe levels.

The dataset of 200 hypoxic fractions was also included in a correlation analysis against HIF1A protein expression assessed by IHC (Fig. 7A). A strong correlation between HIF1A expression and hypoxic fraction was found for a HL_MRI threshold of −0.21 (P = 0.0021; Fig. 7B) within the interval for mild hypoxia. This HL_MRI value was related to the cell line–derived hypoxia activation level of 2.2% O₂ (Fig. 5A), which is comparable with findings for HIF1A stabilization in experimental studies (8, 9). Moreover, this HL_MRI of −0.21 was outside the range for which a significant association to treatment outcome was found (Fig. 5C), consistent with results from survival analysis based on HIF1A protein (Fig. 7C). Taken together, by our imaging method, it appeared possible to distinguish hypoxia levels with association to different biological processes like cancer hallmarks and HIF1A stabilization.

We present a method based on diagnostic MRI to visualize hypoxia levels in patient tumors. Previous imaging studies, including our development of the CSH-tool, have focused solely on the presence of hypoxia without considering its severity (15, 16). By utilizing the CSH-tool together with a dimensional reduction algorithm to assign weighted information of oxygen consumption and supply to each individual pixel, totally new information reflecting the continuous range of hypoxia levels in tumors, was obtained. Although adding more factors like cellular proliferation rate or blood oxyhemoglobin saturation into our algorithm may improve the technology, comparison of our results with indirect measures of hypoxia levels by pimonidazole staining and gene expression showed strong correlations and validated the method. The MRI-defined hypoxia levels differed in their association to treatment outcome and cancer hallmarks in cervical cancer, demonstrating that new understanding of how various levels affect tumor aggressiveness and biology can be achieved. Our approach is easily applicable in the hospital's diagnostic procedures, and a step toward a better exploitation of MR images in the clinic.

Our algorithm was validated in xenograft tumors by matched comparison with pimonidazole-based images as independent surrogate measures of hypoxia levels. Pimonidazole has been considered as a classical IHC marker for hypoxia and used for validation in numerous imaging studies (1). We observed a steady decrease in staining intensity away from necrosis, consistent with work by others (26, 27). In the same manne, binding of pimonidazole or other nitroimidazole compounds in cells or pieces of tissue cultured in vitro has been shown to decrease with increasing oxygen concentrations (24, 28, 29). It is therefore likely that the intensity gradients in our histologic sections reflected true differences in hypoxia levels. By using large-scale similarity analysis of hypoxic fractions obtained from MRI and pimonidazole staining, followed by correlation analysis of the corresponding levels, the linear range for reliable detection of hypoxia levels in xenograft tumors was obtained. The algorithm was further confirmed in patient tumors by using an indicator of hypoxia levels based on gene expression. Strict criteria for gene selection, using expression responses in two cell lines exposed to a range of oxygen concentrations and correlation analysis of expression and imaging data in patient tumors, revealed nine suitable indicator genes. Indeed, a strong linear relationship between the cell line–derived hypoxia activation levels and HL_MRI was found, confirming that a continuous range of hypoxia levels could be visualized in patient tumors.

Caution should be taken to directly transfer the oxygen concentrations for activation of genes in cell lines to hypoxia levels in patient tumors. However, it would enable a rough comparison of our results with existing pO₂ data of cervical cancer. The median MRI-defined hypoxia level for all tumors combined corresponded to a cell line–derived level of 1.3% O₂, which is within the range of 3–17 mmHg (∼0.4%–2.2%) achieved by oxygen electrodes (14). Moreover, the strongest correlation to treatment outcome was found for a level corresponding to 0.7% O₂, based on cell line data. This is highly consistent with most pO₂ studies, reporting association to outcome for hypoxic fraction below 5 mmHg (approximately 0.7% O₂; ref. 14). Although surrogate measures of hypoxia levels were achieved by our approach, the levels seemed to be in accordance with oxygen electrode measurements, and to distinguish levels shown to be of prognostic significance in previous work.

Hypoxia levels associated with biological processes like cancer hallmarks and stabilization of HIF1A protein were identified by our method. At moderate hypoxia, that is, the level most strongly correlated with treatment outcome, oxidative phosphorylation, targets of the MYC oncogene, and G₂–M checkpoint, appeared significant, consistent with previous work where we identified a treatment resistant cervix tumor phenotype associated with the same hallmarks (30). This tumor phenotype also appeared to have increased mitochondrial and proliferative activity (30). The stronger correlation of moderate hypoxia levels with poor outcome in our present work might therefore be explained because hypoxic cells still have enough oxygen to proliferate under these conditions, in accordance with a hypothesis proposed by others (10, 31). HIF1A protein expression, on the other hand, appeared significant at mild hypoxia levels, in line with previous reports (6), and showed no correlation to outcome.

At severe hypoxia, the DNA repair hallmark appeared significant, consistent with studies showing activation of DNA damage response at extremely low oxygen concentrations (32). Our finding that inflammatory response and EMT were associated with such severe levels, is less well documented. It is tempting to speculate that this could be a consequence of lactate accumulation due to near complete vascular shut down in regions with severe hypoxia. Lactate is a key molecule in the inflammatory immune suppressive response in tumors (33, 34), and such inflammatory environment is a strong inducer of EMT (35). Although these novel associations for the most severe levels need to be explored further, the findings demonstrate a potential of our method to achieve better insight into the hypoxic tumor phenotype.

Our method to visualize hypoxia levels proposes a new application of routinously acquired DCE-MR images that may have implications for the diagnostic evaluation of patients. The finding that the CSH-tool could be used for this purpose, broadens the utility of the tool. DCE-MRI is widely used for diagnosis of cervical cancer, however, investigation of our approach using other modalities including DW-MRI would be of high interest. Moreover, our work encourages exploration of hypoxia levels in other cancer types as well, by exploiting the MR technology already available at most hospitals. The method thus provides a well-needed opportunity to investigate the importance of individual hypoxia levels in tumor progression.

H. Lyng reports grants from The South-Eastern Norway Regional Health Authority (grant no. 2015020) and grants from The Norwegian Cancer Society (grant nos. 107438 and 182451) during the conduct of the study. No potential conflicts of interest were disclosed by the other authors.

T. Hillestad: Conceptualization, data curation, software, formal analysis, validation, investigation, visualization, methodology, writing-original draft, writing-review and editing. T. Hompland: Conceptualization, data curation, software, formal analysis, supervision, validation, investigation, visualization, methodology, writing-original draft, writing-review and editing. C.S. Fjeldbo: Data curation, formal analysis, investigation, visualization, writing-original draft, writing-review and editing. V.E. Skingen: Data curation, software, formal analysis, investigation, visualization, methodology, writing-original draft, writing-review and editing. U.B. Salberg: Data curation, formal analysis, investigation, writing-original draft, writing-review and editing. E.-K. Aarnes: Data curation, writing-review and editing. A. Nilsen: Data curation, writing-review and editing. K.V. Lund: Data curation, writing-review and editing. T.S. Evensen: Data curation, formal analysis, investigation, writing-review and editing. G.B. Kristensen: Resources, data curation, writing-review and editing. T. Stokke: Resources, writing-review and editing. H. Lyng: Conceptualization, resources, supervision, funding acquisition, writing-original draft, project administration, writing-review and editing.

Financial support was received from The South-Eastern Norway Regional Health Authority (grant number 2015020) and The Norwegian Cancer Society (grant numbers 107438 and 182451).Technical assistance from D. Trinh, Department of Pathology at Oslo University Hospital, is highly appreciated.

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.