Mounting evidence has revealed that genetic and phenotypic variations exist between tumors and within each of the individual tumors (1–3). These variations result in considerable inter-tumoral and intra-tumoral heterogeneities of dose–response to radiotherapy, significantly impacting patient clinical outcomes (4–8). Therefore, targeting individual tumor heterogeneity of dose–response using a spatially non-uniform treatment dose distribution has been suggested and clinically feasible to personalize radiotherapy treatment and improve patient therapeutic ratio (9–14).

Tumor treatment response to radiation is influenced by many biological factors and changes in the tumor microenvironment (15). Most of these factors are unknown before treatment and modified dynamically during the treatment course. Tumor radiosensitivity has been estimated before treatment using in vitro clonogenic assay (16–18) or a linear regression model derived from the specific gene expressions (19–21). However, these methods can only measure the tumor intrinsic cellular radiosensitivity and could not be utilized to assess intra-tumoral treatment dose–response modified by tumor cell repopulation (22–24), reoxygenation (25–28), reactivation of immune response (29, 30), etc. There have been different methods to assess intra-tumoral treatment dose–response at the tumor voxel level using biological imaging feedback, i.e., acquiring PET or MR images during the treatment course (8, 14, 31–37).

Treatment feedback images have the potential to explore dynamic features of cellular activities in tumors during the treatment course, which could guide us to select the most efficient and reliable time points to quantify and estimate treatment dose–response for clinical therapeutic decisions. Most importantly, quantified intra-tumoral dose–response will guide the design of heterogeneity treatment doses to maximize the therapeutic ratio (14). In this study, serial weekly fluoro-2-deoxyglucose (FDG)-PET/CT feedback images were used to evaluate the dynamic characteristics of tumor voxel treatment response with respect to different treatment dose levels. The predictive capability of tumor voxel treatment dose–response was studied to determine the time points and minimal numbers of imaging feedback.

Figure 2 shows the tumor voxel SUV measured during the pretreatment and different treatment weeks during radiotherapy. The SUV declined noticeably from 6.76 ± 3.69 measured at pretreatment to 3.85 ± 2.22 measured at the 2nd treatment week. After that, the SUV continuously declined and became stabilized with a mean ± SD of 3.27 ± 2.25, 2.5 ± 1.79, and 2.23 ± 1.27 for the 3rd, 4th, and last treatment weeks, respectively.

The mean ± SD of the DSC between the contours generated by the DIR and the one manually delineated by the physicians was 0.84 ± 0.06 on the week 2 feedback image and was 0.79 ± 0.08 on the week 4 feedback image. The mean surface distances were (1.66 ± 0.54) mm and (1.85 ± 0.63) mm for the week 2 and the week 4 feedback images. No negative Jacobian value was observed for all tumor voxels. Figure 3 shows the mean ± SD of the Jacobian values for individual patients. The Jacobian value was calculated from the DVFs obtained from the DIR performed between the pretreatment image and the week 2 and week 4 feedback images.

A strong linear relationship between the logarithm tumor voxel SUV change ratio measured within the first 4 treatment weeks at the different treatment dose levels was identified. The mean ± SD of Pearson’s correlation coefficient was 0.91 ± 0.15 for all the 20,757 tumor voxels. As a comparison, the correlation coefficient between the tumor voxel SUV change ratio and treatment dose was lower (p-value <0.001, two-sample t-test) with the coefficient being 0.89 ± 0.18 for all tumor voxels. Figure 4 shows the measured logarithm SUV change ratios versus treatment dose and the linear model (red line) from Eq. 1 for those tumor voxels with the average dose–response DRM being 0.2, 0.4, 0.6, and 0.8, where the average DRM was determined using the first 4 weekly feedback images.

Figure 5 shows the relationship between tumor voxel DRM and tumor voxel SUV change ratio measured at different dose levels derived based on Eqs. 1 and 2. The slope of the curve decreased with the increase of the dose level. It implies that a DRM estimated at early treatment will be more sensitive to the errors in SUV measurement. Therefore, later DRM estimation will be more reliable.

Figures 6A–C shows the cumulative histograms of DRMs for all tumors, control tumors, and failure tumors, respectively. The distributions of DRM_i for both control and failure groups converge gradually to DRM_L. Both control and failure groups have a certain number of resistant tumor voxels (e.g., DRM > 0.8), and their number gradually reduced during treatment due to reoxygenation. However, the percentage of the reduction for the failures remains smaller compared to the controls. For the controls, the percentage of tumor voxels with DRM > 0.8, V(DRM > 0.8), was 30.5%, 16%, 9.5%, and 3.6% for DRM₁, DRM₂, DRM₃, and DRM₄, respectively. For the failure patients, the corresponding V(DRM > 0.8) was 55.8%, 13.4%, 13.4%, and 7.7%. Figures 6C–E show the histogram of the deviation of the DRM, i.e., DRM_i − DRM_L (i = 1, 2, 3, and 4), for different tumor groups. For the control patients, the mean ± SD of DRM deviation was 1.22 ± 3.67, −0.01 ± 0.34, −0.04 ± 0.17, and −0.01 ± 0.08 for the DRM₁, DRM₂, DRM₃, and DRM₄, respectively. For the failure patients, the corresponding DRM deviation was 0.71 ± 4.99, −0.08 ± 0.28, −0.12 ± 0.17, and −0.14 ± 0.12, indicating the systemic underestimation of tumor voxel resistance.

Figure 7 shows the mean ± SD of the deviation between the DRM_L and each of the DRMs estimated using a single feedback image for the tumor voxels within each level of DRM_L. Figure 7A shows that the DRM₂ has larger deviations as compared to the DRM₃ and DRM₄ for tumor voxels with respect to different levels of DRM_L in absolute value for all tumor voxels. In contrast, Figure 7B shows the relative deviations (%) decrease with the increase of the DRM_L level.

Figures 8A, B show the pretreatment (SUV₀) and ith treatment week (SUV_i) for a local failure tumor. Figure 8C shows the 6-month posttreatment (post-Tx) FDG-PET/CT image. The locally high metabolic activity region (arrow) detected the recurrence position. Figure 8D shows the tumor voxel DRM estimated using a single weekly feedback image. The highly resistant areas (DRM > 1) on the DRMs, predicted using different weekly PET feedback images, appeared to be consistently close to the recurrence location.

Table 2 shows the ROC test results for the DRMs estimated using different PET feedback images. For those DRMs estimated using a single feedback image, the predictive capability quantified by AUC was improved from 0.78 for DRM₁ to 1.00 for DRM₄. In contrast, the predictive capability of the DRMs estimated using 2 feedback images remains high, with AUC ≥ 0.95. The predictive capability of the FDG avidity or SUV_max remained relatively low. The AUCs for the maximum SUV were 0.61~0.77. Both MTV and TLG achieved a moderate predictive capability with the AUC being 0.74~0.81 and 0.77~0.93, respectively. There is no clear time trend for the predictive capability of SUV_max, MTV, and TLG. Tables 3 – 5 summarize the details of the ROC analysis for the SUV_max, MTV, and TLG measured at the pretreatment and different treatment weeks.

By utilizing tumor voxel SUV change ratio determined using serial FDG-PET/CT imaging feedback, tumor voxel dose–response can be predicted and combined with the pretreatment SUV to guide treatment dose adaptation (14). The previous study (8) has demonstrated that tumor voxel DRM assessed for head and neck squamous cell carcinoma (HNSCC) had very large intra- and inter-tumoral variations. The variation had similar numerical distribution to the variations of cellular intrinsic radiosensitivity index or in vitro SF₂. In addition, the DRM index was found to strongly correlate with the expression of cancer stem cell biomarker CD44 for HNSCC patients (48). Tumor voxel DRM is a dynamic index that is constantly modified by the delivered radiation dose during treatment. Thus, DRMs, estimated using imaging feedback acquired at different treatment dose levels, could have different values reflecting the dynamic characteristic of radiation-induced tumor voxel dose–response. The current study demonstrated that the tumor voxel DRM became relatively stabilized after the 2nd treatment week ( Figure 2 ) or the dose > 20 Gy (for 2 Gy per fraction treatment regimen). However, the stability was dependent on the DRM levels. For those of more resistant tumor voxels, i.e., DRM > 0.8, a larger variation could occur in the later treatment ( Figure 7 ). Tumor voxel DRM estimated using either 1 or 2 feedback images acquired within the dose range of 30~40 Gy can predict tumor voxel dose–response and be used for treatment adaptation.

Tumor voxel DRM estimated using a single feedback image is most favorable in clinical practice. The estimated DRM was sensitive to the timing of the PET image feedback. The predictive capability of the DRM improved as the treatment dose level increased ( Table 2 ). The feedback image acquired in the 1st treatment week or within the dose range of (7.4 ± 1.8) Gy had minimal predictive capability. The disadvantage of using a very early FDG-PET feedback imaging in tumor response assessment has been studied. A study reported that using the feedback image acquired at the 1st treatment week cannot predict histomorphological tumor response (49). Another study reported that there was an initial increase of FDG uptake in tumors after 6 Gy of radiation dose (50). These very early (dose ¾ 10 Gy) metabolic changes could be caused by the inflammatory and immune response in tumors, which obscures the changes in tumor glucose metabolism induced by therapeutic effects (50). The feedback image acquired at the 2nd treatment week had an improved predictive capability and reduced systematic deviation with respect to the DRM_L constructed in the latest treatment week. However, DRM estimated using the 2nd feedback image still has a larger deviation from the DRM_L compared to the one obtained at the 3rd or 4th treatment week ( Figure 7 ). The single feedback image obtained within the 3rd or 4th treatment week was comparable with respect to the DRM_L ( Figure 6 ). However, each of them has different advantages and disadvantages. Feedback image acquired within the 3rd treatment week will provide an early prediction and, thus, more room for clinical treatment adaptation. However, an earlier DRM estimation is more sensitive to the uncertainty in the DRM construction ( Figure 5 ). Using Eqs. 1 and 2, one can derive the relationship between the uncertainty of tumor voxel SUV and DRM, as follows:

Thus, the uncertainty in DRM construction caused by the SUV uncertainty is inversely proportional to the treatment dose d. As an example, 5% of SUV variation will cause about 7.9% of DRM variation predicted at a dose of 20 Gy, compared to 5.3% at a dose of 30 Gy. Therefore, the time point of single imaging feedback acquired after the 2nd treatment week or >20 Gy faces a tradeoff between the early clinical decision on treatment adaptation and predictive reliability. One clinical option is to select the single feedback time point based on the minimal treatment dose required in a clinical dose painting protocol.

The predictive capability of tumor voxel DRM on treatment outcome of local–regional tumor failure or control was slightly improved when using 2 PET feedback images in the DRM estimation ( Table 2 ). Meanwhile, the predictive capability was less sensitive to the timing of feedback imaging. Therefore, if the clinical workload is not a major concern, using 2 feedback images acquired between the end of the 2nd and 4th treatment weeks should be favorable to be used as guidance for target dose adaptation.

Various markers have been developed to predict tumor response to radiation for HNSCC patients. A gene expression profiling created from biopsies was proved to be of high predictive value on treatment outcomes (51, 52). Specific expression patterns of microRNA have been shown to predict therapeutic response in HNSCC patients (53, 54). In addition, as shown in this study and others (55–57), the SUV_max, MTV, and TLG obtained from a single FDG-PET image also had good predictive value on the treatment outcome. These markers are all useful in predicting treatment outcomes. However, the purpose of the study is not to demonstrate that the DRM can provide a better or equal prediction of treatment outcome than the other markers. The predictive capability quantified in the study was for evaluating the time point of image feedback for DRM construction. In fact, we used both the pretreatment SUV and DRM obtained from image feedback together to assess treatment outcome (14). Using both the pretreatment tumor voxel SUV, as a surrogate of tumor cell density in tumor voxel, and the tumor voxel DRM, as the radiosensitivity, to assess treatment outcome quantitatively also provides the spatial distribution of tumor voxel dose-efficacy. This important 3D information in tumors will be used to optimize dose distribution design for individual patients in adaptive treatment.

The major weakness of the study was that the patient cohort, especially the failure patient number, is relatively small. Therefore, the predictive capability of the patient outcome should be further validated by a larger patient cohort. To achieve a target power of 0.95, a future clinical trial will need at least 49 patients assuming the null hypothesis to be AUC = 0.5. Here, the target power was defined as the desired probability of rejecting a false null hypothesis. The patient number was estimated using a previously published method (58). Another weakness was the limited number of PET/CT feedback images for each patient used in the study. Due to different clinical reasons and because a number of patients missed 2 to 3 weekly feedback images required in the protocol, the dynamic characteristics of DRM could not be reliably explored.

Tumor voxel metabolic activity measured using FDG-PET/CT feedback images fluctuated noticeably during the first 2 treatment weeks and then became stabilized thereafter. Single FDG-PET/CT imaging acquired after the 2nd treatment week or the treatment dose >20 Gy is recommended to predict tumor voxel dose–response matrix in the current clinical practice. The time point of image feedback can be selected based on clinical application; later time points should be more reliable.

The data that supports these findings are available on request from the corresponding author.

The studies involving human participants were reviewed and approved by Beaumont Health Institutional Review Board. The patients/participants provided their written informed consent to participate in this study.

SC performed the data acquisition and analysis and wrote the manuscript. AQ developed the deformable image registration code and evaluated the image registration. DY designed the study, supervised the research project, and revised the manuscript. All authors listed have made a substantial, direct, and intellectual contribution to the work and approved it for publication.

SC has been supported by Beaumont Health in this study.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.