Data

For this study, which was approved by the Institutional Review Board, FDG PET/CT scans from 58 subjects diagnosed with HNSCC from 2004–2008 were available for retrospective data analysis. This population was selected because of the availability of both imaging and long-term clinical follow-up data, which enables assessing the prognostic characteristics of features. Descriptive statistics of the data are presented in Table 1. We note that while standard clinical features like age, gender, smoking status were available for the studied subjects, they did not appear to add predictive power to the radiomics features presented. Furthermore, image data as well as corresponding clinical metadata (sex, age, smoking status, drinking history, stage, primary site location, etc.) are part of a head and neck cancer data collection available on NCI’s The Cancer Imaging Archive in DICOM format [5, 6]. In addition, it is noteworthy that the patients were largely accrued prior to the recognition of HPV as an important prognostic variable and hence this laboratory value was not available for the vast majority of the cohort. Furthermore, the majority of cancers were treated with primary chemo-radiotherapy. Chemotherapy was platinum based and given either weekly or on an every three week concomitant schedule. Primary cancers included the following anatomical regions: base of tongue, oropharynx, pyriform sinus, tonsil, hypopharynx, and nasopharynx. The cases had varying T and N stage, but no distant metastases. Two FDG PET/CT scans were analyzed for each subject—a pre-treatment and a post-treatment scan. Patients were treated after full evaluation at an interdisciplinary tumor board, with most receiving definitive chemoradiotherapy. A PET/CT for response assessment was obtained at 8–12 weeks after completion of radiation therapy [7, 8]. For imaging, the clinical standard protocol was used and performed on Siemens Biograph 40, Siemens Biograph Duo, or GE Medical Systems Discovery LS PET/CT scanners. All subjects were injected with 370 MBq ± 10% of [F-18]FDG with an uptake time of 90 minutes ± 10%. In all cases subjects were fasted for >4 hours and had blood glucose <200 mg/dL. Because of the interest in the H&N region, patients were imaged with arms down, and CT-based attenuation correction was performed. All reconstructions were performed with iterative 2D OSEM algorithms. The voxel size ranged from 3.4×3.4×2.0 to 4.3×4.3×5.0 mm, with the majority at 3.5×3.5×3.4 mm.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Demographic and clinical characteristics.

https://doi.org/10.1371/journal.pone.0215465.t001

Calculation of quantitative imaging features

First, all FDG PET volumes were SUV normalized by utilizing the PET DICOM Extension [9] for 3D Slicer, a multi-platform free and open software package for visualization and medical image computing [10]. Second, an experienced radiation oncologist inspected all scans and identified primary tumors, which were segmented by utilizing the semi-automated segmentation approach described by Beichel et al [4]. The approach utilizes a graph-based optimization algorithm for segmentation, requires little user interaction, and is available for 3D Slicer in the form of an editor effect extension [11]. All segmentations were performed following standard clinical practice in radiation oncology. Note that out of 58 cases, 25 showed residual uptake in the primary lesion on the 8–12 week follow-up PET scan. These lesions were also segmented using the same segmentation tool. Consequently, no segmentation was performed for 33 follow-up PET scans due to complete response after treatment. Third, the five most commonly utilized basic and 17 new features were calculated utilizing the PET-IndiC extension for 3D Slicer [12]. An overview and detailed description of all calculated features is given in Table 2. The majority of new features were designed to characterize the SUV uptake pattern within the segmented lesion by utilizing descriptive statistics. One exception is the feature SAM (standardized added metabolic activity), which was proposed by Mertens et al. [13] with the goal to develop a partial volume independent marker of total lesion glycolysis. SAM utilizes a small rim region around the segmented lesion for partial volume correction. Note that Mertens et al. [13] proposes two versions: SAM and normSAM. In this work we utilize SAM, and similarly as most of the other features listed in Table 2, we have normalized this feature by the uptake in the liver as described below in more detail. Furthermore, we have included the average SUV of this rim region as a separate feature, which was dubbed RA. An example for the rim around a lesion is given in Fig 1. Fourth, to account for different metabolic base-line activity, we further normalized SUV measurements to the metabolic activity of the liver at the time of imaging, similarly as proposed by Wahl et al.[2]. Such an approach is especially important for change assessment using pre- and post-treatment PET scans. For normalization to metabolic baseline activity, a spherical liver measurement region was automatically defined by utilizing the approach described by Bauer et al.[14], which is also available in 3D Slicer in the form of an extension [15]. From this liver measurement region, the average SUV is calculated, and all SUV-based features were normalized by dividing them by this uptake value. For cases with pre-treatment (t₀) and post-treatment (t₁) segmentations, the change of the quantitative image features was calculated as .

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1.

Example of a rim region (green mask) used for calculating feature RA relative to the lesion segmentation (red mask).

https://doi.org/10.1371/journal.pone.0215465.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Overview of image-derived features utilized.

https://doi.org/10.1371/journal.pone.0215465.t002

Statistical analysis

Survival data analysis methods were used to assess the effect of each feature on DFS. DFS was defined as time from initial treatment to disease recurrence or death. Patients for whom events were not observed during their follow-up were treated as censored observations at the date of last follow-up in the analysis. Descriptive survival probabilities were estimated and plotted with the method of Kaplan-Meier for patients grouped into high, intermediate, and low imaging feature categories according to tertiles. Quantitative imaging feature effects on survival were modeled individually with Cox regression. Regression estimates are reported as hazard ratios (HRs) for one standard deviation increases in the imaging feature along with 95% confidence intervals (CIs). Rank-order agreement (concordance) between prognostic scores from the regression models and observed survival times was estimated with the c index [16]. This index can be interpreted as the probability that, of two randomly selected patients, the one with a better prognostic score will survive longer. Values of 0 or 1 indicate perfect concordance and 0.5 no concordance. Two-sided p-values for tests of the significance of features in the regression models are reported, unadjusted for the number of tests performed. To account for multiple statistical testing, false-discovery rate (FDR) was also computed with the method of Benjamini and Hochberg [17]. FDR can be used to select among the full set of features tested so as to control the expected proportion of false positive statistical findings. Features were identified as significant in this study at a FDR of 10%. Statistical analysis was performed with the survival [18] and survminer [19] packages in version 3.4.1 of the R statistical software [20].

Survival associations and impact

Results of pre-treatment FDG PET scans confirm the findings reported by Castelli et al.[1]; MTV and TLG perform better than max and peak markers. In this context, note that the number of papers utilizing the peak marker included in the review performed by Castelli et al. [1] is limited. In our study, we opted to normalize tracer uptake by the average liver uptake. Omitting this step would lead to a similar ranking (MTV: p = 0.0283, TLG: p = 0.1464, Peak: p = 0.8292, and Max: p = 0.9764). A potential reason for the better relative performance of MTV and TLG could be that MTV and TLG are spatially more “inclusive” than max and peak measurements, accounting for the full volumetric and metabolic extent of the tumor, rather than a simple assessment of the highest level of metabolic activity. The new features RA and Glycolysis Q2 showed similar performance to MTV and TLG on pre-treatment scans. Because multiple statistical testing is employed, we report p-values adjusted to control the false-discovery rate (FDR), and a FDR of 10% was specified in this study for the determination of statistical significance. Based on this criterion, none of the features at baseline is deemed to be significant due to the number of features investigated combined with the limited dataset size.

When looking at the survival effect of post-treatment feature value change in those selected cases where primary lesions were still visible in the follow-up FDG PET scan, it is important to seek potential imaging based biomarkers that might distinguish the high proportion of false positive scans from those indicating true recurrence. In our analysis, RA is ranked highest (FDR: 0.048) and found to be significant at a 10% FDR. This is a particularly challenging group since those with residual uptake have approximately a 50% chance of having residual disease (true positive) versus 50% with no true recurrence that are false positives. In this setting, MTV and TLG were not found to statistically distinguish between true and false positives.

We speculate that this finding can be explained as follows. The region of interest for RA is defined as the two voxels wide rim surrounding the segmented tumor volume. Hence, RA captures the metabolic activity in the primary tumor’s periphery (Fig 1). Typically, cases with recurrence had higher RA-values in the pre-treatment scan compared to the post-treatment scans (positive ), indicating a loss of activity in this rim region potentially due to low or reduced immunogenicity in the tumor. Patients with false positive scans tended to have a lower RA to start and tended to see an increase in RA with treatment, which appeared to portend a better outcome (negative ) perhaps due to enhanced inflammation or immunogenicity in this region. The suggestion of less inflammation with recurrence highlights the complex interplay between radiation, chemotherapy and tumor. A summary of RA pre- and post- treatment trend patterns is provided in Fig 3. This suggests that an increase in activity in this peripheral region is predictive for improved outcome and we hypothesize this is consistent with recently supported data showing that increased immune response to tumors portends better outcomes [22–25]. Furthermore, it may help differentiate between true positive recurrences from false positive results that have been a common issue in post-treatment PET-imaging [7, 26–31]. Similarly, closely nearby-ranked features Minimum and 1st Quartile uptake, also capture aspects of the tracer uptake in the region of transition from hot lesion core to background, especially in the case of non-necrotic lesions, which represents the vast majority of cases investigated in this work (Fig 1). This is also reflected in the correlation of features Minimum and 1^st Quartile with RA (0.8022 and 0.8496, respectively). In this context, there are several issues to consider. First, out of these three features, only RA has a FDR below 10%. Second, the feature Minimum is quite sensitive to image noise when compared to RA, and thus, using Minimum is not recommended. Third, while there is room for optimization of features like RA and 1st Quartile their results point in the same direction. Furthermore, standard features (e.g., max, peak, MTV, or TLG) were not significantly associated with survival when looking at post-treatment feature change.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Summary of RA values at time points t₀ (pre-treatment) and t₁ (post-treatment).

(a) Boxplots. (b) False positive cases and (c) true positive cases. The dashed line indicates the average trend.

https://doi.org/10.1371/journal.pone.0215465.g003

When considering both pre-treatment and change at post-treatment, RA is the only feature that has a p-value<0.05 before FDR correction. In addition, when correcting for FDR, it stays significant as a post-treatment change based feature. Therefore, we conclude that RA represents a good candidate for further investigations. Clearly, RA is functionally the reverse of what is currently broadly utilized by the community (e.g., SUV_max, MTV, TLG) as it does not focus on the core region of the lesion, but instead on the periphery. In this context, one exception is the feature SAM, which corrects for partial volume effects by utilizing the tumor rim region [13].

Based on our findings, we conducted a literature search to see whether similar features were investigated for the purpose of outcome prediction. We found one recent publication that describes a tumor shell-based radiomics approach, which analyzes the tumor periphery [32]. The shell feature captures a sequence of morphologic patterns across the primary tumor boundary, such as shape, size, SUV values, and heterogeneities in a simple 2D map, which is vectorized to form a feature vector and is subsequently used as the input for a support vector machine (SVM) classifier. In [32], the authors demonstrated the ability of this approach in predicting treatment response based on pre-treatment PET scans for patients receiving stereotactic body radiation therapy (SBRT) for early stage non-small cell lung cancer and for patients receiving external beam radiation therapy (EBRT) and concurrent chemotherapy followed by high-dose-rate intracavitary brachytherapy (ICBT) in stage IB-IVA cervix cancer. While the work reported in [32] uses a different approach for deriving tumor periphery features and assesses them in two different applications, we conclude that the lesion rim is promising for future investigations. Furthermore, the increased activity in the periphery of lesions could represent enhanced immune response, which could have potential utility in determining activity of immune therapy interventions. So far, assessing therapy response from these agents has been difficult and hence the impact could be substantial, but requires significant investigation. In addition, future histologic evaluation of these rim regions is needed to confirm or refute this potential explanation.

Future work

The goal of this work was to identify potentially promising quantitative imaging features. As such, the presented work is part of an ongoing evaluation process, following the consensus approach outlined in [21]. In our future work, we will assess the impact of imaging equipment, reconstructions, as well as the acquisition settings on feature performance. Also, some features might be more sensitive to variability in lesion segmentation than others, and the impact of rim thickness on predictive performance will be investigated. In this work, only the primary lesion was considered. However, efficient PET image segmentation tools like the one presented in Beichel et al. [4] make it feasible to individually segment involved lymph nodes. At this stage, it is unclear how to best utilize this additional information and if this could potentially add predictive value. Furthermore, in future work we plan on increasing our evaluation database considerably and studying the impact of different uptake reference regions (aortic arch, cerebellum, etc.) for normalization on prediction performance of features.