Machine learning identifies multi-parametric functional PET/MR imaging cluster to predict radiation resistance in preclinical head and neck cancer models

About 50% of patients treated with radiochemotherapy (RCT) for locally advanced human papilloma virus–negative head-and-neck cancer (HNC) experience local and regional treatment failure [1, 2]. As salvage treatment options are limited, locoregional failure in most patients leads to severe symptoms and ultimately to death. Thus, overcoming treatment resistance by optimized RCT represents an important area of research. Preclinical and clinical data demonstrates that tumor hypoxia and other microenvironmental factors significantly contribute to tumor radiation resistance [3‐6]. Different quantitative imaging biomarkers (QIBs) related to tumor hypoxia and microenvironment have shown potential for outcome prediction, early response assessment, and RT personalization, e.g., by means of risk adapted radiation dose modulation [7‐12].

Hypoxia imaging using positron emission tomography (PET) with specific radiotracers such as [18F]-Fluoromisonidazole (FMISO) has proven prognostic power to predict outcome after RCT in HNC [7, 13‐15]. Similarly, functional magnetic resonance imaging (MRI) techniques, such as diffusion-weighted (DW) imaging assessing tumor cellularity or dynamic contrast-enhanced (DCE) imaging which allows to analyze tissue vascularity and vessel permeability, have been correlated to tumor response after RCT in HNC and other solid tumors [8, 9, 16, 17]. Some studies correlated the spatial distribution of multiple QIB and suggested complementary biological information [18‐20]. However, the optimal QIB or imaging profile using multiple QIB to predict outcome after RCT in HNC is unknown. Most results were derived from small observational clinical cohorts and none of the previous studies was able to relate relevant QIB to radiation resistance on a biological or pre-clinical level.

Future clinical use of QIB to personalize radiation dose to overcome treatment resistance requires a widely available, robust, affordable, and simple method to generate QIB to allow multicenter trials and easy access for patients. In contrast to molecular profiling [21, 22], liquid biopsy [23, 24], histopathology [25, 26], or combination with immunotherapy [27, 28], QIBs have the benefit of spatial tumor characterization [29] and thus optimal conditions for focal personalized interventions such as dose-painting, including dose escalation and dose de-escalation [13, 30, 31].

The aim of this preclinical study was to develop and train a multi-scale model from a broad and unbiased basis for prediction of high-risk subvolumes (HRS) in HNC linked to increased radiation resistance derived from hypoxia PET, DW-, and DCE-MRI. Multi-parametric small animal PET/MRI of xenograft tumors from different human HNC cell lines with variable, known radiation sensitivities were imaged and evaluated by novel machine learning (ML) methods to identify HRS in multi-dimensional imaging space. The hypothesis to be investigated in this study was therefore that with novel ML approaches new QIB or imaging profiles will be discovered to define HRS in a pre-clinical scenario, which may be used for future personalized radiotherapy (RT) interventions in a clinical setting.

During model training in 1D to 5D search space on the baseline imaging data, we identified distinct clusters in 1D to 3D imaging parameter space which were able to significantly stratify the xenograft tumors according to their radiation resistance. Using 1D parameters only, the best stratifying cluster was obtained for ADC derived from DW-MRI which showed the highest stratification power with an effect size [95% CI] of \({\mathrm{S}}_{\mathrm{HRS},1\mathrm{D}}=1.69 \left[1.46-3.20\right]\), \(p=0.002\). Interestingly, the center (interval) of the cluster was located at \({\mathrm X}_{1\mathrm D,\mathrm{ADC}}=420\left[384;457\right]\cdot10^{-6}\text{mm}^2\text{/s}\), which corresponded to the left flank of the histogram generated from all animals. In comparison, we found significantly increased stratification potential for a 2D cluster defined by ADC and FMISO_c1 (\({\mathrm{S}}_{\mathrm{HRS},2\mathrm{D}}=2.68 \left[2.41-4.12\right]\), \(p=0.01\)). Details on cluster center and interval in terms of imaging parameter values are given in Table 3. The best stratifying cluster in n-dimensional imaging space was spanned by the 3D quantitative maps of ADC, FMISO_c1, and FMISO_c2, yielding an effect size of \({\mathrm{S}}_{\mathrm{HRS},3\mathrm{D}}=2.99 \left[2.50-4.44\right]\), with \(p<0.0001\) respectively. When further increasing the dimensionality of the parameter space, further improvement of \({\mathrm{S}}_{\mathrm{HRS}}\) was observed, which, however, was not significant (\(p>0.05\)) according to Mann-Whitney U test based on a bootstrap analysis with respect to \({\mathrm{S}}_{\mathrm{HRS},3\mathrm{D}}\). Best scoring models in 1D to 5D imaging space are summarized in detail in Table 3. A visualization of the n-dimensional search space is presented in Fig. 4, whereas Fig. 5 shows the corresponding stratification potential for the selected 1D, 2D, and 3D clusters. Figure 6 presents an example of one preclinical tumor (SAS) with annotations of 1D and 3D HRS.

Correlation of cell line specific radiation sensitivities with the classical imaging parameters in the tumor region did only yield significant stratification potential for \({ADC}_{valley}\) (\(p=0.006\)), cf. Figure 7 and Table 4.

Figure 7 shows the validation results for the best 1D, 2D, and 3D models identified during training in addition to the only significant classical parameter \({ADC}_{valley}\). Stratification results of the different models in the extended cohorts \({C}_{max}\) are similar to those obtained the training cohort \({C}_{all}\), indicating high robustness of the method.

Following the same methodology, 1D to 5D parameter space scanning was performed for imaging data obtained after 2 weeks of fractionated RT. Here, only a 1D cluster defined by the FMISO_c1 map measured in w2 yielded significant stratification potential \({\mathrm{S}}_{\mathrm{HRS},1\mathrm{D},\mathrm{ w}2}=1.12 \left[0.90-3.69\right]\), \(p=0.041\). Results of n-dimensional model training in w2 of RT are detailed in Table 5.

In this study, we report pre-clinical training of a multi-dimensional PET/MRI-based QIB to detect HRS in HNC as potential target for future focal dose escalation. Our findings suggest that a HRS defined by a cluster of ADC values derived from DW-MRI correlates spatial maps of cellularity with individual radiation resistance considering a 1D quantitative functional imaging map as input. Highest stratification potential with respect to cell line specific radiation resistance was found for a 3D QIB created from ADC, and two PCs of dynamic FMISO PET information. Increasing dimensionality further did not significantly increase stratification potential, which may be due to redundancies hidden in the n-dimensional functional imaging data. Consequently, we identified a QIB profile from PET/MRI using a novel machine learning approach in a pre-clinical setting. Starting from a wide search approach with as few assumptions as possible using the main quantitative imaging techniques which are clinically available today, we were able to identify the most promising multi-parametric QIB for potential usage for future RT individualization.

The proposed method relies on the identification of a radioresistant cluster in parameter space only. Consequently, we do not per se assume a spatially connected area of the HRS inside the tumor. If spatial connection is given, HRS may be used for potential future local radiotherapy interventions, such as dose painting. If HRS voxels in contrast would be scattered throughout the tumor, this might be indicative of a generally more radioresistant tumor and dose painting strategies may result in a radiation dose escalation of the whole tumor. However, scattered HRS voxels throughout the GTV might also be caused by noise and potentially weak robustness of the model, which should be clarified in future validation studies in preclinical and ultimately also clinical settings.

Due to their limited size and heterogeneity, direct application of the ML models to identify spatially connected HRS regions in patients may not be possible. In this study, eight different cell lines with distinct radiation resistance levels were used, meaning that each small animal tumor must be understood as a role model for one voxel of a patient tumor. Consequently, the final model may not necessarily yield connected HRS areas but will require retraining and validation in patients.

ADC has been identified by earlier studies as potential prognostic QIB in HNC [8, 16], whereas other studies reported controversial results [37]. The discrepancy of earlier results may be due to over-simplified imaging measures such as mean ADC averaged over the whole tumor in contrast to the sub-volume approach based on clusters in multi-dimensional QIB space proposed in this study. Classical or global imaging parameters investigated in this study demonstrated that \({ADC}_{valley}\) appears to also be associated with radiation sensitivity. A potential explanation for this observation might be that \({ADC}_{valley}\) is a mean value calculated from seven voxels around the minimum \(ADC\) in a tumor sample and may thus be correlated to the 1D cluster identified during ML training on voxel level.

However, when using joint information from ADC maps derived from DW-MRI combined with two PC of dynamic FMISO PET, significantly better stratification was obtained compared to ADC only. This comes however to the expense of acquiring in addition to DW-MRI dynamic hypoxia PET which increases the level of complexity during patient examination and image acquisition enormously. So far, only small hypoxia PET patient data sets were reported due to the complexity of acquisition requiring experimental tracer production, extended scan times, and non-standard data analysis strategies which make a broad roll-out of this technology unrealistic [5]. Nevertheless, these findings corroborate earlier results reported by our group and others that dynamic hypoxia PET has prognostic character with respect to RCT outcome [7, 12, 15]. Assuming that repeated functional imaging will further enhance the power of image-based adaptive RT interventions, it appears that dynamic hypoxia PET is more complex, costly, and not as broadly available as DW-MRI. Thus, from a pragmatic point of view, DW-MRI appears promising for wider clinical roll-out with change of practice even if less predictive than 3D-HRS combining DW-MRI and FMISO PET.

Analysis of the preclinical imaging data acquired 2 weeks after fractionated RT revealed no stratification of radiation resistance groups for most cluster combinations. Sole hypoxia PET yielded slightly significant stratification power at this time point early during RT. As such, this confirms clinical findings of prognostic potential of FMISO PET at the second week during RT [14, 38]. However, in this study, the model for w2 was newly trained without any inference from the models obtained for pre-treatment data.

Our ML approach used to identify multi-dimensional clusters of radiation resistance is based on several assumptions. First, radiation resistance levels were based on data from earlier pre-clinical studies [32, 33], showing significant variation in radiation resistance between experiments. Second, small animal functional imaging is extremely challenging, requires anesthetized animals, and thus deviates from a standard clinical situation. In addition, we assumed a relative HRS size varying between 0 and 20% depending on the radio-resistance levels of the respective cell lines. A further drawback of our method is the fact that parameter space scanning was performed directly on image voxel data, which is more prone to noise and registration inaccuracies compared to volume averaged methods. An alternative would be to combine single voxels to small homogeneous subregions (supervoxels) prior to parameter space scanning, e.g., by means of simple linear iterative clustering [19].

In this study, we used a data-driven ML approach in terms of PCA for extracting a reduced number of QIB maps from dynamic functional imaging. The use of PCA for dynamic data has been shown to be promising by other clinical and pre-clinical studies [39] providing potentially more robust results compared to classical use of compartment models for such data [7, 9].

A previous study proposed deriving high-risk tumor subvolumes from joined functional imaging information by clustering patient imaging data [19]. However, this method does not directly use the size of a HRS for patient stratification but apply different intermediate steps to determine heuristic stratification parameters. In contrast, our method uses the relative HRS size which is directly connected to cell line specific hypoxia levels which are only available in a translational approach. This prior represents a major limitation of our study, as no tumor specific hypoxia or radiation resistance levels were measured. This underlines the necessity of independent validation studies, ideally in patients to confirm the hypotheses identified in this experiment.

Potential uncertainties of the method making use of multi-dimensional functional imaging data on voxel level originate from manual contouring of tumor regions used as input for the analysis as well as co-registration of the functional imaging data sets which is of crucial importance for the integrity of the data set in higher dimensions. Robustness of the proposed HRS method was evaluated in different ways. The density of visualized scores in parameter space (Fig. 4) shows a smooth distribution as well as a single, compact region of high scores \({S}_{HRS}\), indicating robust learning of the cluster center \({X}_{HRS}\), which is further supported by the internal bootstrap validation using the training cohort \({C}_{all}\). Furthermore, robustness of the model was evaluated using an extended cohort \({C}_{max}\) including additional tumors which were not part of the initial training cohort \({C}_{all}\). Even though this evaluation indicated stability of the model parameters, this approach cannot be considered a full independent validation due to only a small number of additional data sets in \({C}_{max}\) compared to \({C}_{all}\). A potential alternative for tumor stratification based on joint QIB maps might be an end-to-end learning approach using for example convolutional neural networks (CNNs), which have shown to achieve high performance in image processing and classification tasks [40]. We did not investigate such approach since we had only a low number of tumors with the full multi-dimensional imaging parameter space available in this study (n = 42). Therefore, an approach was developed which complements a data-driven learning method with hypotheses about the existence and size of an HRS related to known radio-resistance levels. The final model can easily be interpreted in the sense that learned HRS are fully determined by associated QIB ranges. In contrast, model interpretation using CNN-based end-to-end learning might be challenging.

In Fig. 5, cell line UTSCC-45 shows distinctly different HRS compared to all other cell lines of the group with high radiation sensitivity (group H). Interestingly, this cell line differs from the other investigated cell lines due to its positive human papilloma virus (HPV) status. The associated genetic difference may cause a shift in radiosensitivity compared to HPV-negative cancer cell lines which seems not to be detectable by quantitative imaging [41]. Therefore, ADC/FMISO-based HRS radiation dose escalation does not seem an option for low-risk HPV-positive oropharyngeal HNC and future interventional trials should be limited to patients with high-risk profiles (HPV-negative or HPV-positive plus  > 20 pack-years smoking history) [42].

As tumor hypoxia and cellularity are subject to change during RCT, individualized RT approaches adapted to the current level of resistance will only be possible if HRS can be identified shortly before treatment. Recently developed hybrid MR-Linacs may allow functional MRI acquisitions before and during RT and open thus unique possibilities in terms of MR-specific QIB-adaptive RT [43]. Recent results on phantom and early clinical data proved that quantitative imaging is possible at hybrid MR-Linac systems [44, 45] which is a major pre-requisite for biologically adapted RT dose painting based on ADC clusters. More complex multi-parametric QIB involving different imaging modalities may need to be acquired on dedicated PET/MRI scanners and used for offline response-adaptive RT. Nevertheless, before QIB-based RT dose painting can be applied in clinical RT practice, technical and clinical validation is required including test–retest studies and comparison to diagnostic scanners to ensure repeatability and reproducibility [43, 46].