Predictive modeling of outcomes following definitive chemoradiotherapy for oropharyngeal cancer based on FDG-PET image characteristics

From 12/2002 to 3/2009, all stage III–IV OC patients treated at our institution with definitive concurrent chemoradiation therapy with retrievable pre- and post-treatment FDG-PET/CT scans were identified. To ensure relative uniformity in technique and quality of scans for subsequent analysis, any patient with an FDG-PET/CT scan performed outside of our institution was excluded unless a repeat FDG-PET/CT scan was obtained at our institution prior to initiation of treatment. Patients with metastatic disease at presentation, noted on staging or treatment planning imaging, were also excluded as were those patients who were managed surgically. This resulted in 174 assessable patients. Our Institutional Review Board approved this retrospective study and all patients provided informed consent.

Pre-treatment FDG-PET imaging was performed as part of the standard treatment planning process for definitive concurrent chemoradiation therapy. The mean uptake time was 72  ±  17 min. Procedures for FDG-PET/CT imaging at our institution have been previously described (Lim et al 2012, Romesser et al 2014). In brief, patients are instructed to fast for a minimum of 6 h, with water intake permitted and encouraged. Prior to administration of 18F-FDG (dose range, 12–15 mCi), a blood glucose level of  <200 mg dl⁻¹ is confirmed. 18F-FDG is injected intravenously, followed by an uptake period during which patients drink diluted oral contrast. Low-dose CT (120–140 kV, 80 mA) and PET scans are then obtained for the torso (3 min/bed position, thoracic inlet to upper thigh) with the arms up, followed by dedicated images of the head and neck (5 min/bed position) with the arms down. Intravenous contrast is also administered for radiotherapy planning scans. PET images were acquired on either a hybrid PET/CT GE or Siemens system, normalized and corrected for scatter, randoms, attenuation, decay and dead time, and reconstructed using the ordered-subsets expectation maximization (OSEM) algorithm (2 iterations and 8 subsets for Siemens scanner, and 2 iterations for GE scanners with 28, 21, and 20 subsets for Discovery LS, ST, and STE, respectively).

Pre-treatment attenuation-corrected FDG-PET scan images were converted to the Computational Environment for Radiotherapy Research (CERR) format. CERR is an open source radiotherapy research toolkit designed to facilitate developing and sharing research results for radiotherapy planning; it is MATLAB-based software and provides a common file type for the creation of multi-institutional treatment plan databases for various types of research studies, including dose-volume outcomes analyses and radiomics studies (Deasy et al 2003). Bounding boxes were generated for the primary lesion by creating cuboidal structures that encompassed the individual FDG-avid elements. For the purposes of this study, SUV was defined as the decay-corrected measurement of activity per unit volume of tissue (MBq ml⁻¹) adjusted by the total administered activity (MBq) and divided by the patient's weight (kg) measured on the date of the scan. A threshold of 42% of the maximum SUV value (SUV_max) was then applied to define a region of interest (ROI) for further analysis, and the volume of ROI was defined as MTV (Erdi et al 1997).

A total of 24 representative features of the FDG-avid regions, defined as ROIs, were then extracted for each image. These features include common statistical features such as the SUV_max and quantizations of the intensity-volume histogram distribution of SUV values over the defined ROI volume. Additionally, more complex shape and textural features that take morphological features and second-order gray-level co-occurrence matrix (GLCM)-based features of the analyzed ROI into account were included (Leijenaar et al 2013, Huang et al 2016, Lian et al 2016, Sutton et al 2016). These are further described in the following sections.

First-order statistical features were derived from the distribution of voxel values over the analyzed ROI intensity-volume histogram. They include kurtosis, skewness, slope, and the minimum, maximum, median, and average value of the SUV. Kurtosis is a measure of the flatness or 'peakedness' of the intensity-volume histogram, whereas skewness is a measure of the asymmetry of the intensity-volume histogram. Slope is the change in volume over the change in the SUV threshold used to generate the ROI volume.

Shape features are related to the morphology of the ROI itself and in this study include eccentricity, solidity, extent, and Euler number. Eccentricity is a measure of 'non-circularity' defined as the ratio of the minor axis to the major axis of the best fitted ellipsoid to the analyzed ROI, with an eccentricity of 0 or 1 corresponding to a linear or a perfectly round ROI, respectively. Solidity is derived by calculating the proportion of pixels of the ROI to the largest possible convex hull polygon structure of the ROI; the convex hull is the best-fitting polygon that encloses all the pixels of the ROI. An ROI of the same shape and volume as the convex hull would have a solidity of 1 whereas an irregular ROI would have solidity  <1. Extent is similar to solidity except that a rectangular prism or cuboid is used, instead of a convex hull, to measure the proportion. The Euler number is an integral value that indicates the number of connected objects in the ROI minus the number of holes.

The GLCM was constructed by summing up the co-occurrence frequencies for each matrix of 13 directions across a 3D image with 16 gray-scale levels. Texture features quantify the voxel-to-voxel interaction within the ROI, and include homogeneity, entropy, contrast, and coherence (also known as energy in Haralick texture features) (Haralick et al 1973, Tesar et al 2008). These features are independent of the ROI position, orientation, and size and reflect the distribution of tumor metabolic uptake while minimizing the potential contribution of variations in administered dose and time to FDG-PET image acquisition (El Naqa et al 2009). Homogeneity is a measurement of the similarity in intensity of each voxel and its neighboring voxels whereas contrast is a measurement of the variability for the difference in intensity between each voxel and its neighbors. Entropy is a measurement of the randomness of the intensity level over the voxels in the ROI, and coherence measures the uniformity of the intensity level in the ROI. As these texture features are dependent on the intensity relationships between neighboring voxels, noise or artifacts caused in the image acquisition process may greatly impact on the values of imaging features. To remove or minimize noise or artifacts, a smoothing method was applied: for a voxel (v) in the ROI, a sphere with a user-defined radius (from 0.5 cm to 1.5 cm with a step of 0.5 cm) with the voxel (v) being centered was created. An intensity value averaged from all voxels in the sphere was placed in the voxel (v). This procedure was performed for all voxels in the ROI. Image features were extracted from the smoothed image. The post-smoothing was applied after reconstruction.

All outcomes were measured from the start of radiation therapy to the time of event. All-cause mortality (ACM) time was defined as the time to death from any cause. LF time was defined as the time to any LF in the high-dose region of radiation therapy treatment. DM time was defined as the time to first clinical or pathological evidence of distant disease recurrence. Patients were censored at the date of last follow-up if death did not occur.

Univariate and multivariate survival analyses were performed using Cox proportional-hazards regression for ACM whereas competing-risks analysis, based on Fine and Gray's proportional sub-hazards model, was used for LF and DM (Fine and Gray 1999). After univariate analysis on clinical variables, multivariate analysis was performed with variables with p-values  <  0.05. For each endpoint, Kaplan–Meier analysis with log-rank test was performed to investigate the difference of survival rates between a lower risk group and a higher risk group (Wilson 1927, Kaplan and Meier 1958). Statistical analysis was performed using Stata/MP version 12 (Stata Corporation, College Station, TX).

Multiparameter logistic regression models were built incorporating clinical factors and imaging features. For the best-fitting model selection, leave-one-out cross validation (LOOCV) with forward feature selection was performed. A model that occurs with the most frequency during the LOOCV process was chosen as a final model for each endpoint. Models were characterized by the area under the receiver operating characteristic (ROC) curve (AUC) and p-values were computed using the Spearman's correlation coefficient (Hanley and McNeil 1982, Borkowf 2000).

For an unbiased model estimate, a 5 fold cross validation method was iterated 30 times (Myles et al 1997). At each iteration, univariate logistic regression analysis was performed, and features with p-values  <  0.05 were used in multiparameter logistic regression as shown in the above section, resulting in an AUC computed between the predicted and original outcomes. After the whole process, AUC values were averaged. In addition, ROC curve analysis was performed to determine an optimal cutoff value using Youden's index based on which sensitivity and specificity were computed.

Additionally, final outcome models (as shown in the above section) were tested on an independent cohort, consisting of 65 patients with stage III–IV OC treated with IMRT-based chemoradiation therapy at another institution (Washington University School of Medicine, St. Louis, MO, USA) from 7/2003 to 11/2009. PET images were acquired on a Siemens Biograph Duo scanner or a Siemens Biograph 40 scanner. Patients were instructed to fast for a minimum of 4 h. Prior to administration of 18F-FDG (dose range, 10–15 mCi), a blood glucose level of  <200 mg dl⁻¹ was confirmed. A spiral CT scan was obtained at approximately 60 min postinjection and noncontrast CT images were obtained for attenuation correction and fusion with PET images. After that, emission images were obtained. For segmentation, a threshold of 42% of the maximum SUV value was used, and the same set of image features used in this study was extracted. For more information about this dataset, see Garsa et al (2013).

We analyzed 174 HNSCC patients with stage III–IV oropharyngeal cancer following definitive chemoradiation therapy who were treated at our institution. Among them, 48 (27.6%) patients died, and 12 (6.9%) and 33 (19.0%) patients had LF and DM, respectively. Characteristics of the primary study cohort are provided in table 1. Median follow-up time was 55 months (range: 6–112 months). The majority (87.4%) of the patients were male, and most (69%) were current or former smokers (20.7% and 48.3%, respectively). The average age was 57 years (range: 27–84 years). The OC subsite was the tonsils in 47.1% of patients, base of the tongue in 48.9% of patients, and the soft palate or posterior pharyngeal wall in 4% of patients. The overall stage was IV in 78.7% of patients and III in 21.3% of patients; 39.1% of patients had T3 or T4 disease, and the majority had N2 disease (68.4%); only a small portion of patients in our cohort presented with N3 disease (4.6%). The median value for MTV was 11.2 cc (range: 2.2–60.9 cc).

All (100%) patients were treated with definitive concurrent chemoradiation therapy. The median dose to the tumor was 70 Gy (range: 66–70 Gy), and the median dose to the lower neck was 50.4 Gy (range: 50–70 Gy). Concurrent chemotherapy consisted of cisplatin alone in 56.1% of patients, cetuximab alone in 10.4% of patients, carboplatin and 5-fluorouracil in 12.1% of patients, cisplatin and bevacizumab in 13.3% of patients, and other multidrug regimens in the remaining 8.1% of patients.

For pre-treatment scans, the median uptake time was 67 min. To investigate whether the variability in uptake time has impact on the results, patients were split into two groups with a cutoff of 67 in uptake time and four texture features were compared using Wilcoxon rank-sum test. No significant differences were found between the two groups with coherence (p  =  0.396), contrast (p  =  0.880), entropy (p  =  0.445), and homogeneity (p  =  0.855).

The four most common sizes of voxel were 3.91  ×  3.91  ×  4.25 mm (n  =  56; 32.2%), 4.69  ×  4.69  ×  3.27 mm (n  =  46; 26.4%), 5.33  ×  5.33  ×  4.00 mm (n  =  28; 16.1%), and 5.15  ×  5.15  ×  2.40 mm (n  =  20; 11.5%). Using a Kruskal–Wallis test, the comparison of texture features between patients with the four different voxel sizes resulted in non-significance for coherence (p  =  0.278), contrast (p  =  0.707), entropy (p  =  0.547), and homogeneity (p  =  0.778).

We performed a Kruskal–Wallis test to investigate whether there are significant differences in texture features between different PET/CT scanners. No significant differences were found with coherence (p  =  0.271), contrast (p  =  0.783), entropy (p  =  0.479), and homogeneity (p  =  0.855). These results imply that the impact of variability in uptake time, voxel size, and scanner on texture features is not significant in this cohort.

For each endpoint, the predictive power of individual clinical factors (T, N, overall stage, smoking status, location, Karnofsky performance status (KPS), age, sex, and biologically equivalent dose with alpha/beta  =  10 [BED₁₀]) and extracted imaging features were assessed using logistic regression. When a smoothing sphere with a radius of 0.5 cm was used, better power was achieved than that without the smoothing method. For instance, contrast and homogeneity features showed statistical significance in LF with AUC  =  0.69 (p  =  0.035) and AUC  =  0.70 (p  =  0.026), respectively, using the smoothing method, whereas there was no significant texture feature without the smoothing method. Thus, for all tests, we used imaging features extracted after smoothing the ROI with a sphere with a radius of 0.5 cm. Table 2 shows single factor AUC and p-value results that had statistical significance (p  <  0.05) for at least one endpoint, resulting from univariate logistic regression analysis. For comparison, SUV_max and SUV_mean were also displayed, which did not show statistical significance for all three endpoints. KPS, T stage, extent, skewness, MTV, and TLG had significant correlations with all three endpoints. The best predictors were T stage (AUC  =  0.67, p  <  0.001) and skewness (AUC  =  0.67, p  <  0.001) for ACM whereas MTV was significantly associated with both LF and DM with AUC  =  0.81 (p  =  0.001) and 0.66 (p  =  0.004), respectively. BED₁₀ did not reach statistical significance for all three endpoints with AUC  =  0.49 (p  =  0.799), 0.51 (p  =  0.731), and 0.51 (p  =  0.490) for ACM, LF, and DM, respectively.

For the multiparameter logistic regression models, ACM was correlated to kurtosis and MTV; LF was correlated to homogeneity and MTV; DM was correlated to solidity, kurtosis, and MTV. The models are given in table 3. The 5 fold cross validation used for an unbiased model estimate resulted in AUC  =  0.65 (standard deviation (SD)  =  0.02; p  =  0.004), 0.73 (SD  =  0.04; p  =  0.026), and 0.66 (SD  =  0.04; p  =  0.015) for ACM, LF, and DM, respectively, showing statistical significance for all three endpoints.

Characteristics of the independent validation cohort are provided in table 4 (Garsa et al 2013). All patients in the independent cohort had stage III–IV OC and were treated with definitive concurrent chemotherapy. Those patients who were also surgically managed were excluded, resulting in 65 evaluable patients. Among them, 31 (47.7%) patients died, and 10 (15.4%) and 11 (16.9%) had LF and DM, respectively. Similar to the primary study cohort, the majority were male (78.5%), and the average age was 58 years (range: 38–78 years). Smoking status was unknown in 50.8% of patients, but among those whose information was available, 81.3% were identified as smokers. The majority of patients were stage IV (86.2%); many had advanced primary tumors, with 69.2% presenting with T3 (20.0%) or T4 (49.2%) disease. Most (63.1%) patients presented with N2 disease although there was a larger proportion of patients in the independent cohort with N3 disease (7.7%) than the primary study cohort (4.6%). The median value for MTV was 18.7 cc (range: 3.5–64.7 cc). The models for ACM, LF, and DM were tested on this independent cohort. As shown in table 3, significant predictive power was retained in LF with AUC  =  0.68 (p  =  0.029) whereas the models for ACM and DM showed borderline significance with AUC  =  0.60 (p  =  0.092) and 0.65 (p  =  0.062), respectively.

For ACM, Cox proportional-hazards regression was performed whereas for LF and DM, Fine and Gray's proportional sub-hazards models were performed with each single clinical variable, and, using variables with p  <  0.05, multivariate models were tested. Smoking history (69% were long-time smokers) was not statistically significant. KPS was most commonly associated with outcomes (p  <  0.05 for ACM, LF, and DM) as shown in table 5. Other clinical factors significantly associated with outcomes included T stage with DM and ACM (p  =  0.001 and 0.010, respectively) and N stage with DM (p  =  0.001).

It should be noted that MTV was chosen in all three models as shown in table 3. To investigate the significance of MTV further, patients were split into two groups by median MTV for each endpoint, and a Kaplan–Meier analysis with log-rank test was performed. Statistically significant differences were found for all three endpoints with p  =  0.034, 0.025, and 0.026 for ACM, LF, and DM, respectively. Additionally, patients' MTV values were sorted in ascending order and grouped into three groups of equal size; those patients in the middle group were removed. In the comparison between one-third of patients with smaller MTV and one-third of patients with larger MTV, statistically significant differences were found for all three endpoints with p  =  0.014, 0.006, and 0.037 for ACM, LF, and DM, respectively (figure 1(A)). Similarly, for each endpoint, predicted outcomes obtained using the predictive models shown in table 3 were sorted, and one-third of patients in the middle were removed. In the comparison between the riskiest one-third of patients and the least risky one-third of patients, statistically significant differences were found for all three endpoints with p  <  0.001, 0.006, and  <0.001 for ACM, LF, and DM, respectively (figure 1(B)).

Zoom In Zoom Out Reset image size