Predicting hypoxia status using a combination of contrast-enhanced computed tomography and [18F]-Fluorodeoxyglucose positron emission tomography radiomics features

Abstract

Background and purpose

Hypoxia is a known prognostic factor in head and neck cancer. Hypoxia imaging PET radiotracers such as ¹⁸F-FMISO are promising but not widely available. The aim of this study was therefore to design a surrogate for ¹⁸F-FMISO TBR_max based on ¹⁸F-FDG PET and contrast-enhanced CT radiomics features, and to study its performance in the context of hypoxia-based patient stratification.

Methods

121 lesions from 75 head and neck cancer patients were used in the analysis. Patients received pre-treatment ¹⁸F-FDG and ¹⁸F-FMISO PET/CT scans. 79 lesions were used to train a cross-validated LASSO regression model based on radiomics features, while the remaining 42 were held out as an internal test subset.

Results

In the training subset, the highest AUC (0.873 ± 0.008) was obtained from a signature combining CT and ¹⁸F-FDG PET features. The best performance on the unseen test subset was also obtained from the combined signature, with an AUC of 0.833, while the model based on the 90th percentile of ¹⁸F-FDG uptake had a test AUC of 0.756.

Conclusion

A radiomics signature built from ¹⁸F-FDG PET and contrast-enhanced CT features correlates with ¹⁸F-FMISO TBR_max in head and neck cancer patients, providing significantly better performance with respect to models based on ¹⁸F-FDG PET only. Such a biomarker could potentially be useful to personalize head and neck cancer treatment at centers for which dedicated hypoxia imaging PET radiotracers are unavailable.

1. Introduction

Head and neck squamous cell carcinomas (HNSCC) treated with chemoradiotherapy can reach overall survival rates of nearly 90% when human papillomavirus (HPV)-positive (HPV⁺) tumors dominate the cohort [1, 2, 3]. However, head and neck tumors with regions of insufficient oxygenation –a condition known as hypoxia– are still associated with poor prognosis [4]. Studies have shown that hypoxia triggers angiogenesis, increases radioresistance and reduces the effectiveness of surgery, an alleged consequence of hypoxia selected pressure for tumor aggressiveness and metastasis [5, 6, 7, 8]. This has motivated the development of HNSCC chemoradiotherapy protocols that stratify patients based on the assessment of tumor hypoxia [9, 10, 11]. Major alterations of the treatment are being tested in clinical trials, such as the reduction of the radiation dose to non-hypoxic lesions from the standard 70 Gy to 30 Gy, thereby significantly reducing the risk of normal tissue toxicity [12, 13].

The preferred technique to detect hypoxia in this context is currently positron emission tomography (PET) combined with a hypoxia radiotracer such as ¹⁸F–Fluoromisonidazole (¹⁸F-FMISO) [14, 15, 16]. Previous studies have shown that ¹⁸F-FMISO PET is a prognostic biomarker both before and during treatment [17, 18, 19]. However, ¹⁸F-FMISO is still an investigational agent that requires an Institutional Review Board (IRB) with an Investigational New Drug approval from the FDA. The task of extending existing hypoxia-based HNSCC trials to the routine clinical setting would be simpler if an equivalent biomarker could be obtained from standard imaging techniques.

Hypoxia has been shown to increase ¹⁸F-Fluorodeoxyglucose (¹⁸F-FDG) uptake in cancer cell lines [20]. This is not however the only contribution to the ¹⁸F-FDG signal, and studies assessing the overall correlation between ¹⁸F-FDG and the level of hypoxia obtain conflicting results [21, 22]. In this study we sought radiomics features to explore the hypothesis that an image-based biomarker obtained from the combination of ¹⁸F-FDG PET scans and their accompanying contrast-enhanced CT scans correlates with ¹⁸F-FMISO-assessed hypoxia, with contrast-enhanced CT scans acting as surrogates for blood flow [23]. In particular, we analyze its utility for the task of discriminating with enough specificity a subset of candidate patients for dose de-escalation, using only imaging techniques employed routinely in the clinical setting. To maximize the information extracted, we quantify the images using radiomics features [24] and develop the biomarker using machine learning.

2. Materials and methods

2.1. Dataset and hypoxia definition

Patients were treated for head and neck cancer with chemoradiation at Memorial Sloan Kettering Cancer Center (MSKCC) under a de-escalation trial (IRB #04-070; NTC00606294 on clinicaltrials.gov). To be included in the present retrospective analysis, patients had to have received pre-treatment ¹⁸F-FDG static PET/CT for treatment planning, as well as ¹⁸F-FMISO dynamic PET/CT. Scans were performed in a GE Medical Systems PET/CT scanner with the patient immobilized with a customized mask. Prior to the ¹⁸F-FDG PET/CT scan patients were given 100 ml of omnipaque contrast. The ¹⁸F-FMISO PET/CT scan consisted of a 30-minute dynamic scan followed by two 10-minute static scans obtained approximately 90 and 150 minutes post-injection, respectively. The average time between the two PET scans was 6 days, with a standard deviation of 5 days. All patients with the available data treated between January 2011 and February 2016 were included in the analysis.

The lesions were delineated by experienced physicians on the contrast-enhanced CT for radiotherapy treatment planning. ¹⁸F-FDG uptake was quantified in terms of the body-weight standardized uptake value (SUV). ¹⁸F-FMISO uptake was normalized by the uptake measured in the ipsilateral (with respect to the lesion) jugular vein. The level of hypoxia of a lesion was defined in terms of its maximum tumor-to-blood uptake ratio (TBR_max) on the last static scan. In particular, in this study a lesion was considered to be hypoxic if TBR_max > 1.4¹.

2.2. Radiomics features

Radiomics features were extracted from both ¹⁸F-FDG PET and CT scans, including itensity features (from PET and CT), shape features (CT), and texture features (CT). Intensity features also included variables based on two-dimensional volume-intensity histograms, for example the volume spanned by voxels with intensities above 70 Hounsfield units (HU), denoted V_>70. Features were extracted in Matlab using the CERR Radiomics Imaging Quantification toolbox (RIQ, [25, 26], April 2017 version). All code used to derive features is open source and available for download. Table S1 in the Supplementary Material includes a full list of the radiomics features computed.

In CT scans, radiomics features were extracted for each lesion from four different volumes of interest, denoted v_GTV (gross tumor volume, GTV), v_CT∼ (GTV after filtering out voxels outside of HU ∈ [−100, 150]), v_↑ (SUV > 42% SUV_max) and v_↓ (SUV < 42% SUV_max), respectively, as shown in Figure 1.

Figure 1.

Feature extraction pipeline, indicating the volumes used to compute each family of features. IVH = Intensity Volume Histogram, RLM=Run Length Matrix, NGTDM=Neighborhood Gray-Tone Difference Matrix, Neighboring Gray-Level Dependence Matrix.

For ¹⁸F-FDG PET features, the volume of interest was defined as the region within the GTV with SUV > 42% SUV_max [27]. We use the intersection between the fixed-threshold contour and the GTV to avoid the overestimation of small lesion boundaries [28].

2.3. Data analysis

All lesions with a volume larger than 10 cm³ were considered for the analysis. The dataset was divided into a training subset comprising approximately 65% of the lesions, and an internal test subset which has held out and only used to test the final models. TBR_max was used as the continuous response variable to predict with a supervised learning model. However, as the ultimate goal was stratification, the response was dichotomized by classifying lesions with TBR_max > 1.4 as hypoxic, and the performance was evaluated in terms of the area under the receiver operating characteristic curve (AUC).

First, the training dataset was tested for any univariable associations between clinical predictors and the lesions TBR_max. Correlations between numerical predictors and TBR_max were measured in terms of the Spearman correlation coefficient, while associations with categorical predictors were assessed using balanced one-way ANOVA. p-values were corrected for multiple-testing using the Benjamini-Hochberg procedure.

Figure 2 illustrates the radiomics modelling process, which was performed in steps: (1) feature selection based on cross-validated LASSO regression; (2) multivariable linear model building, based on the selected features; and (3) evaluation of the best linear model on the hold-out test set. The process was performed independently for CT and ¹⁸F-FDG PET features. Steps (2) and (3) were also performed for the combined set of selected PET and CT features.

Figure 2.

Flow chart of the steps followed to derive the predictive model, including (1) feature selection, (2) model building, and (3) testing. The curved, dashed line indicates the addition of interaction terms between the selected features to assess the importance of non-linearity.

The feature selection step is based on a LASSO linear model embedded in a 10-fold cross-validation loop reshuffled 10 times. In each fold features were pre-selected by applying a univariable p-value cut and removing those that were highly correlated. The remaining features were used in a LASSO linear model with nested 5-fold cross-validation to determine the regularization parameter yielding the smallest mean square error. The selected features were used again as input to a second LASSO model, this time also including bilinear interaction terms, to assess the non-linearity of the problem. The whole process, outlined in Figure 2, was repeated multiple times to scan over possible values of the pre-selection cuts (p = 0.01, 0.02,… 0.1, Spearman r = 0.5, 0.6… 0.8) and find the ones that maximized the cross-validation AUC.

For the model building step, only features selected in 50% of the 100 cross-validation runs or more were used. Multiple 1- and 2-variable linear regression models were created by taking all possible combinations of the selected features. An optimistic bound on the expected performance of the models was determined in terms of the mean AUC obtained from 10-fold cross-validation reshuffled 10 times.

For each category (PET, CT, or PET+CT) only the linear model with the best AUC was evaluated on the test dataset. The final model coefficients were determined by fitting to the entire training subset.

To assess whether the test AUCs of the three models were significantly better than a model based only on $P_{90\mathrm{\%}}^{\mathrm{FDG}}$ (the 90th percentile of the ¹⁸F-FDG SUV, used here as a robust variant of the maximum SUV), we computed 1000 bootstrap replicas of the test dataset, calculated the corresponding AUCs, and derived a p-value based on the two-sided Wilcoxon rank sum test.

Further details about the definition of the radiomics features and the data analysis procedure can be found in the supplementary material.

3. Results

75 patients satisfying all the requisites were identified, adding up to 121 lesions in total. A randomly chosen subset of 79 lesions were used for training, and the remaining 42 were held out for testing purposes. Patient characteristics are listed in Table 1. None of them were found to be significantly different between the training and testing datasets.

Table 1.

Characteristics of the lesions used in the analysis. For tumor subsite, HPV and p16 status, p-values are based on the Fisher exact test. For the mean and standard deviations of the ¹⁸F-FDG maximum SUV and ¹⁸F-FMISO TBR_max, p-values are based on the 2-sample t-test.

Type	Ntrain (Fraction)	Ntest (Fraction)	p-value
Subsite
Base of Tongue	40 (0.51)	21 (0.5)	1
Hypopharynx	1 (0.01)	1 (0.02)	1
Supraglottic larynx	2 (0.03)	1 (0.02)	1
Tonsil	32 (0.41)	19 (0.45)	0.7
Unknown	4 (0.05)	0 (0)	0.3
HPV status
HPV+	45 (0.57)	27 (0.64)	0.6
HPV−	11 (0.14)	7 (0.17)	0.8
Unknown	23 (0.29)	8 (0.19)	0.3
p16 status
p16+	57 (0.72)	34 (0.81)	0.4
p16−	7 (0.09)	4 (0.1)	1
Unknown	15 (0.19)	4 (0.1)	0.2
Type	Training Mean (SD)	Test Mean (SD)	p-value
18F-FMISO TBRmax	1.91 (0.73)	1.76 (0.56)	0.3
18F-FDG SUVmax	13.64 (5.76)	13.74 (5.23)	0.9

3.1. Association with clinical variables

TBR_max was tested for association with HPV status, p16² status and tumor subsite, and no significant correlation was found. Clinical variables were therefore not included in the radiomics models.

3.2. Model training

CT model

The optimal pre-selection cuts were p_cut = 0.02 and r_cut = 0.6. Two variables were selected by the LASSO model 50% of the time or more: the long-run high-grey-level emphasis of the v_↑ region, denoted ${\mathcal{E}}_{\uparrow }$; and V_>70 of the v_↓ region, denoted $V_{>70}^{\downarrow }$. The specific definition of ${\mathcal{E}}_{\uparrow }$ that was selected was sensitive to texture directionality; in particular, it used the run-length matrix calculated along the direction that yielded the maximum result in that tumor. The resulting cross-validated AUC was 0.78, as shown in Table 2. Adding interaction terms did not improve the result. Using these two features only in a multivariable regression model yielded a mean cross-validation AUC of 0.841, higher than using any of the two features alone.

Table 2.

Performance of the models derived from the training subset assessed in terms of the mean and standard deviation (SD) of the cross-validation AUC. AUCs are also given for the test subset, indicating the bootstrap p-value of the difference with respect to the model with only $P_{90\mathrm{\%}}^{\mathrm{FDG}}$.

Modality	Feature pre-selection	Interaction terms	AUC
	Training		mean ± SD
CT	None	No	0.78 ± 0.03
	None	Yes	0.78 ± 0.03
	ε↑, $V_{>70}^{\downarrow }$	No	0.841 ± 0.007
PET	None	No	0.852 ± 0.008
	None	Yes	0.853 ± 0.007
	$P_{90\mathrm{\%}}^{\mathrm{FDG}}$	No	0.855 ± 0.005
	$P_{90\mathrm{\%}}^{\mathrm{FDG}}$, ${\stackrel{\sim }{\mu }}_3^{\mathrm{FDG}}$	No	0.861 ± 0.007
PET + CT	$P_{90\mathrm{\%}}^{\mathrm{FDG}}$, ${\mathcal{E}}_{\uparrow }$	No	0.873 ± 0.008
	Test		value (p)
PET	$P_{90\mathrm{\%}}^{\mathrm{FDG}}$	No	0.756
PET	$P_{90\mathrm{\%}}^{\mathrm{FDG}}$, ${\stackrel{\sim }{\mu }}_3^{\mathrm{FDG}}$	No	0.767 (p < 0.001)
CT	ε↑, $V_{>70}^{\downarrow }$	No	0.828 (p < 0.0001)
PET + CT	$P_{90\mathrm{\%}}^{\mathrm{FDG}}$, ${\mathcal{E}}_{\uparrow }$	No	0.833 (p < 0.0001)

PET model

The optimal pre-selection cuts were p_cut = 0.01 and r_cut = 0.5. The maximum SUV, here defined in a robust way via the 90th percentile, $P_{90\mathrm{\%}}^{\mathrm{FDG}}$, was selected 100% of the time. In addition, the skewness of the SUV distribution, ${\stackrel{\sim }{\mu }}_3^{\mathrm{FDG}}$, was selected exactly 50% of the time. The cross-validated AUC was 0.852, with no improvement gained from adding interactions. Using these two features only in a multivariable regression model resulted in a mean cross-validation AUC of 0.861, higher than using any of the two features alone.

Combined CT and PET model

All the possible combinations between the top two CT and PET features were considered as predictors in a multivariable linear regression model. The highest performance was achieved when combining $P_{90\mathrm{\%}}^{\mathrm{FDG}}$ and ${\mathcal{E}}_{\uparrow }$, with an AUC of 0.873. Adding interaction terms did not improve the performance.

All the intermediate performance results can be found in Table I of the supplementary material.

3.3. Model testing

Four multivariable regression models were tested on the unseen test subset: (i) ${\mathcal{E}}_{\uparrow }$ and $V_{>70}^{\downarrow }$, (ii) $P_{90\mathrm{\%}}^{\mathrm{FDG}}$, (iii) $P_{90\mathrm{\%}}^{\mathrm{FDG}}$ and ${\stackrel{\sim }{\mu }}_3^{\mathrm{FDG}}$, (iv) ${\mathcal{E}}_{\uparrow }$ and $P_{90\mathrm{\%}}^{\mathrm{FDG}}$. They were trained on the entire training dataset. The performance obtained in terms of AUC was 0.828, 0.736, 0.767 and 0.833, respectively, as shown in Table 2. The highest performance was achieved by the combined $P_{90\mathrm{\%}}^{\mathrm{FDG}}$ and ${\mathcal{E}}_{\uparrow }$ model, in agreement with what was observed in the training dataset. The discriminative power of the two features can be seen in Figure 3a, while the correlation between the linear combination of the two features and ¹⁸F-FMISO TBR_max is shown in Figure 3b. The AUC of the combined model was significantly higher than the AUC of the model based only on $P_{90\mathrm{\%}}^{\mathrm{FDG}}$ (p < 0.0001).

Figure 3.

(a) Scatter plot of $P_{90\mathrm{\%}}^{\mathrm{FDG}}$ against ${\mathcal{E}}_{\uparrow }$ for the test dataset. The markers are color coded according to their TBR_max value. The grey lines result from applying fixed thresholds to the model derived from the training dataset. The resulting classifications have the following false positive and true positive rates for TBR_max > 1.4: t₁ FPR=0.07, TPR=0.72; t₂ FPR=0.31, TPR=0.76; t₃ FPR=0.46, TPR=0.97. Threshold t₁ would result in 6 normoxic lesions out of 13 being classified as such, with only one hypoxic lesion being misclassified as normoxic. Threshold t₂ would result in 8 normoxic lesions being correctly classified, with 7 hypoxic lesions being incorrectly classified as normoxic. Finally, threshold t₃ would classify 12 out of 13 normoxic lesions correctly, but it would incorrectly classify 9 hypoxic lesions as normoxic. (b) Correlation plot of the combined $P_{90\mathrm{\%}}^{\mathrm{FDG}}$ and ${\mathcal{E}}_{\uparrow }$ signature versus F-FMISO TBR_max in the test dataset (Spearman ρ = 0.57, p < 0.01). For values of the PET+CT predictor below 2.0, the Spearman ρ is 0.72, p < 0.01.

4. Discussion

This study explores the possibility of obtaining a surrogate hypoxia biomarker derived from conventional imaging techniques. The proposed signature is based on ¹⁸F-FDG PET and its companion contrast-enhanced CT scan, and reaches an AUC of 0.83 in the test dataset.

We used TBR_max to quantify hypoxia as it minimizes uptake normalization errors and is highly correlated with the tumor-to-muscle ratio that is commonly used in clinical trials [9, 13, 30]. It has been argued that for a more reliable determination of tumor hypoxia a kinetic analysis of dynamic PET data may be needed [31], as the raw ¹⁸F-FMISO uptake is expected to be affected by factors such as the distribution volume [32]. An interesting follow-up study would be to find predictors of the irreversible binding rate of ¹⁸F-FMISO, as determined from kinetic analysis.

We found that the most predictive ¹⁸F-FDG PET features were the 90th percentile and the skewness of the SUV distribution, denoted $P_{90\mathrm{\%}}^{\mathrm{FDG}}$ and ${\stackrel{\sim }{\mu }}_3^{\mathrm{FDG}}$respectively. ${\stackrel{\sim }{\mu }}_3^{\mathrm{FDG}}$ captures the bias of the distribution towards either the higher or lower ends of the SUV spectrum, and its role could be to capture the fraction of hot (and potentially hypoxic) voxels. While $P_{90\mathrm{\%}}^{\mathrm{FDG}}$ remains important in models including CT features, the contribution of ${\stackrel{\sim }{\mu }}_3^{\mathrm{FDG}}$ becomes negligible.

We studied features extracted from CT scans independently, but the analysis was in fact not completely agnostic to ¹⁸F-FDG PET, as the volumes of interest included high-SUV and low-SUV subregions of the GTV. The two top CT features were the volume taken by voxels with HU> 70 within the low-SUV subvolume (denoted $V_{>70}^{\downarrow }$) and the long-run high-grey-level emphasis calculated along the direction with the maximum value within the high-SUV subvolume (denoted ${\mathcal{E}}_{\uparrow }$). By measuring the volume of high contrast-CT density voxels, $V_{>70}^{\downarrow }$ is adding a blood flow dimension to ¹⁸F-FDG uptake. Visual examination of the lesions shows that those with low $V_{>70}^{\downarrow }$tend to have low SUV, CT-hypodense cores such as the one in Figure 4c; it is therefore identifying lesions as necrotic (negligible ¹⁸F-FMISO TBR_max due to limited blood flow) or viable (non negligible TBR_max). On the other hand, ${\mathcal{E}}_{\uparrow }$ is sensitive to successions of voxels (‘runs’) with high CT density within the high-SUV subregion. This suggests that it could be capturing the presence of blood vessels, as in the examples shown in 4a (dense vasculature) and 4b (single blood vessel). The focus on high-¹⁸F-FDG uptake regions could be helping reveal whether the uptake is due to perfusion or hypoxia.

Figure 4.

Representative slices from the CT and ¹⁸F-FDG PET scans of three lesions with a diverse range of radiomics features and TBR_max. Each subfigure corresponds to a different lesion, with the CT scan on the left panel and the ¹⁸F-FDG PET scan on the right panel. The main features for each lesion are: (a) TBR_max=1.5, $V_{>70}^{\downarrow }$ =35 cc, ${\mathcal{E}}_{\uparrow }$ =9735, $P_{90\mathrm{\%}}^{\mathrm{FDG}}$=18.2, ${\stackrel{\sim }{\mu }}_3^{\mathrm{FDG}}$ =0.48; (b) TBR_max=1.8, $V_{>70}^{\downarrow }$ =6 cc, ${\mathcal{E}}_{\uparrow }$ =10980, $P_{90\mathrm{\%}}^{\mathrm{FDG}}$ =11.5, ${\stackrel{\sim }{\mu }}_3^{\mathrm{FDG}}$ = 0.17; (c) TBR_max=1.0, $V_{>70}^{\downarrow }$=0 cc, ${\mathcal{E}}_{\uparrow }$ =5702, $P_{90\mathrm{\%}}^{\mathrm{FDG}}$=1.9, ${\stackrel{\sim }{\mu }}_3^{\mathrm{FDG}}$ =0.65. The white contour indicates v_↓, and the dark grey contour indicates v_↑.

This is not the first time that CT features have been found to be associated with hypoxia. Studies have shown that hypodense neck lymph nodes in routine contrast-enhanced CT scans have hypoxic conditions [23, 33]. Nodal CT density has also been shown to be a prognostic factor in head and neck cancer [34, 35]. More recently, Panth et al found significant differences in the time evolution of CT radiomics features between mice whose hypoxic fraction was reduced due to induced GADD34 over-expression and a control group [36]. In a study by Ganeshan et al, texture features obtained from contrast-enhanced CT images showed significant correlation with the average intensity of tumor staining with pimonidazole in NSCLC patients [37].

Our results suggest that features extracted from routine contrast-enhanced CT features can significantly improve the ability of ¹⁸F-FDG PET to indicate hypoxia, at least to the extent that it can be measured by ¹⁸F-FMISO TBR_max. A large number of radiomics features were considered in the analysis, to maximize the amount of information extracted from the images. The final model only contains two radiomics features, $V_{>70}^{\downarrow }$ and ${\mathcal{E}}_{\uparrow }$, both of which have plausible interpretations.

As seen in Figure 3b, there are outliers in the prediction. Such deviations are expected, as there are multiple situations that can lead to enhanced ¹⁸F-FDG uptake in the absence of hypoxia, including high levels of metabolism, active inflammation, or proliferation rate. Similarly, it has been shown that variations in the distribution volume of ¹⁸F-FMISO can result in particular values of TBR being associated to different levels of hypoxia [32].

From a clinical standpoint, the operating point of the radiomics biomarker would need to be optimized in order to achieve the desired specificity and senbiomarker, assuming sitivity. Figure 3a shows three possible thresholds for the proposed radiomics biomarker, assuming that lesions with TBR_max > 1.4 are defined as hypoxic. In particular, threshold t₁ has a specificity of 93%, and would therefore satisfy the criterion of having a minimal number of false positives included in the de-escalation candidate subset. It may be possible to improve the performance by deriving a higher-order classifier, as opposed to the linear function proposed here; this would however require a larger dataset to avoid overfitting. Our results are only the first step, and suggest that such larger scale effort would be worthwhile.

Key limitations of this study are that it is single institution –although internal validation was successful–, and that the CT imaging protocol was not designed for close monitoring of contrast delivery. Although we believe the results are likely to have broader validity, a multi-institutional study is needed in order to definitively establish this signature (or one like it) to the point that it could be reliably used for patient management. Lastly, the ¹⁸F-FMISO endpoint itself, TBR_max, could be somewhat noisy, which limits the ability to predict its value. Viewed from this perspective, the correlation achieved is promising.

5. Conclusion

We have identified image features derived from conventional imaging that correlate with the magnitude of hypoxia in head and neck cancer patients. In particular, our results show that a combined radiomics biomarker based on ¹⁸F-FDG PET and contrast-enhanced CT can emulate ¹⁸F-FMISO TBR_max-based stratification with significantly higher accuracy than ¹⁸F-FDG PET alone. After validation on large multi-institutional cohorts, such a biomarker could potentially be useful for head and neck cancer patient stratification in situations where ¹⁸F-FMISO is not available.