Voxel Forecast for Precision Oncology: Predicting Spatially Variant and Multiscale Cancer Therapy Response on Longitudinal Quantitative Molecular Imaging

Dynamic Characteristics and Predictive Capability of Tumor Voxel Dose-Response Assessed Using 18F-FDG PET/CT Imaging Feedback

Purpose

Tumor voxel dose–response matrix (DRM) can be quantified using feedback from serial FDG-PET/CT imaging acquired during radiotherapy. This study investigated the dynamic characteristics and the predictive capability of DRM.

Methods

FDG-PET/CT images were acquired before and weekly during standard chemoradiotherapy with the treatment dose 2 Gy × 35 from 31 head and neck cancer patients. For each patient, deformable image registration was performed between the pretreatment/baseline PET/CT image and each weekly PET/CT image. Tumor voxel DRM was derived using linear regression on the logarithm of the weekly standard uptake value (SUV) ratios for each tumor voxel, such as SUV measured at a dose level normalized to the baseline SUV₀. The dynamic characteristics were evaluated by comparing the DRM_i estimated using a single feedback image acquired at the ith treatment week (i = 1, 2, 3, or 4) to the DRM estimated using the last feedback image for each patient. The predictive capability of the DRM estimated using 1 or 2 feedback images was evaluated using the receiver operating characteristic test with respect to the treatment outcome of tumor local–regional control or failure.

Results

The mean ± SD of tumor voxel SUV measured at the pretreatment and the 1st, 2nd, 3rd, 4th, and last treatment weeks was 6.76 ± 3.69, 5.72 ± 3.43, 3.85 ± 2.22, 3.27 ± 2.25, 2.5 ± 1.79, and 2.23 ± 1.27, respectively. The deviations between the DRM_i estimated using the single feedback image obtained at the ith week and the last feedback image were 0.86 ± 4.87, −0.06 ± 0.3, −0.09 ± 0.17, and −0.09 ± 0.12 for DRM₁, DRM₂, DRM₃, and DRM₄, respectively. The predictive capability of DRM₃ and DRM₄ was significant (p < 0.001). The area under the curve (AUC) was increased with the increase in treatment dose level. The DRMs constructed using the single feedback image achieved an AUC of 0.86~1. The AUC was slightly improved to 0.94~1 for the DRMs estimated using 2 feedback images.

Conclusion

Tumor voxel metabolic activity measured using FDG-PET/CT fluctuated noticeably during the first 2 treatment weeks and obtained a stabilized reduction rate thereafter. Tumor voxel DRM constructed using a single FDG-PET/CT feedback image after the 2nd treatment week (>20 Gy) has a good predictive capability. The predictive capability improved continuously using a later feedback image and marginally improved when two feedback images were applied.

Multitask Learning Radiomics on Longitudinal Imaging to Predict Survival Outcomes following Risk-Adaptive Chemoradiation for Non-Small Cell Lung Cancer

(1) Background: Medical imaging provides quantitative and spatial information to evaluate treatment response in the management of patients with non-small cell lung cancer (NSCLC). High throughput extraction of radiomic features on these images can potentially phenotype tumors non-invasively and support risk stratification based on survival outcome prediction. The prognostic value of radiomics from different imaging modalities and time points prior to and during chemoradiation therapy of NSCLC, relative to conventional imaging biomarker or delta radiomics models, remains uncharacterized. We investigated the utility of multitask learning of multi-time point radiomic features, as opposed to single-task learning, for improving survival outcome prediction relative to conventional clinical imaging feature model benchmarks. (2) Methods: Survival outcomes were prospectively collected for 45 patients with unresectable NSCLC enrolled on the FLARE-RT phase II trial of risk-adaptive chemoradiation and optional consolidation PD-L1 checkpoint blockade (NCT02773238). FDG-PET, CT, and perfusion SPECT imaging pretreatment and week 3 mid-treatment was performed and 110 IBSI-compliant pyradiomics shape-/intensity-/texture-based features from the metabolic tumor volume were extracted. Outcome modeling consisted of a fused Laplacian sparse group LASSO with component-wise gradient boosting survival regression in a multitask learning framework. Testing performance under stratified 10-fold cross-validation was evaluated for multitask learning radiomics of different imaging modalities and time points. Multitask learning models were benchmarked against conventional clinical imaging and delta radiomics models and evaluated with the concordance index (c-index) and index of prediction accuracy (IPA). (3) Results: FDG-PET radiomics had higher prognostic value for overall survival in test folds (c-index 0.71 [0.67, 0.75]) than CT radiomics (c-index 0.64 [0.60, 0.71]) or perfusion SPECT radiomics (c-index 0.60 [0.57, 0.63]). Multitask learning of pre-/mid-treatment FDG-PET radiomics (c-index 0.71 [0.67, 0.75]) outperformed benchmark clinical imaging (c-index 0.65 [0.59, 0.71]) and FDG-PET delta radiomics (c-index 0.52 [0.48, 0.58]) models. Similarly, the IPA for multitask learning FDG-PET radiomics (30%) was higher than clinical imaging (26%) and delta radiomics (15%) models. Radiomics models performed consistently under different voxel resampling conditions. (4) Conclusion: Multitask learning radiomics for outcome modeling provides a clinical decision support platform that leverages longitudinal imaging information. This framework can reveal the relative importance of different imaging modalities and time points when designing risk-adaptive cancer treatment strategies.

Gradient Boosted Trees for Spatial Data and Its Application to Medical Imaging Data

Boosting Trees are one of the most successful statistical learning approaches that involve sequentially growing an ensemble of simple regression trees ("weak learners"). This paper proposes a gradient Boosted Trees algorithm for Spatial Data (Boost-S) with covariate information. Boost-S integrates the spatial correlation into the classical framework of eXtreme Gradient Boosting. Each tree is constructed by solving a regularized optimization problem, where the objective function takes into account the underlying spatial correlation and involves two penalty terms on tree complexity. A computationally-efficient greedy heuristic algorithm is proposed to obtain an ensemble of trees. The proposed Boost-S is applied to the spatially-correlated FDG-PET (fluorodeoxyglucose-positron emission tomography) imaging data collected from clinical trials of cancer chemoradiotherapy. Our numerical investigations successfully demonstrate the advantages of the proposed Boost-S over existing approaches for this particular application.

Reliability of Quantitative 18F-FDG PET/CT Imaging Biomarkers for Classifying Early Response to Chemoradiotherapy in Patients With Locally Advanced Non-Small Cell Lung Cancer

PURPOSE OF THE REPORT

We evaluated the reliability of 18F-FDG PET imaging biomarkers to classify early response status across observers, scanners, and reconstruction algorithms in support of biologically adaptive radiation therapy for locally advanced non-small cell lung cancer.

PATIENTS AND METHODS

Thirty-one patients with unresectable locally advanced non-small cell lung cancer were prospectively enrolled on a phase 2 trial (NCT02773238) and underwent 18F-FDG PET on GE Discovery STE (DSTE) or GE Discovery MI (DMI) PET/CT systems at baseline and during the third week external beam radiation therapy regimens. All PET scans were reconstructed using OSEM; GE-DMI scans were also reconstructed with BSREM-TOF (block sequential regularized expectation maximization reconstruction algorithm incorporating time of flight). Primary tumors were contoured by 3 observers using semiautomatic gradient-based segmentation. SUVmax, SUVmean, SUVpeak, MTV (metabolic tumor volume), and total lesion glycolysis were correlated with midtherapy multidisciplinary clinical response assessment. Dice similarity of contours and response classification areas under the curve were evaluated across observers, scanners, and reconstruction algorithms. LASSO logistic regression models were trained on DSTE PET patient data and independently tested on DMI PET patient data.

RESULTS

Interobserver variability of PET contours was low for both OSEM and BSREM-TOF reconstructions; intraobserver variability between reconstructions was slightly higher. ΔSUVpeak was the most robust response predictor across observers and image reconstructions. LASSO models consistently selected ΔSUVpeak and ΔMTV as response predictors. Response classification models achieved high cross-validated performance on the DSTE cohort and more variable testing performance on the DMI cohort.

CONCLUSIONS

The variability FDG PET lesion contours and imaging biomarkers was relatively low across observers, scanners, and reconstructions. Objective midtreatment PET response assessment may lead to improved precision of biologically adaptive radiation therapy.

Value of PET imaging for radiation therapy

This comprehensive review written by experts in their field gives an overview on the current status of incorporating positron emission tomography (PET) into radiation treatment planning. Moreover, it highlights ongoing studies for treatment individualisation and per-treatment tumour response monitoring for various primary tumours. Novel tracers and image analysis methods are discussed. The authors believe this contribution to be of crucial value for experts in the field as well as for policy makers deciding on the reimbursement of this powerful imaging modality.

Value of PET imaging for radiation therapy

This comprehensive review written by experts in their field gives an overview on the current status of incorporating positron emission tomography (PET) into radiation treatment planning. Moreover, it highlights ongoing studies for treatment individualisation and per-treatment tumour response monitoring for various primary tumours. Novel tracers and image analysis methods are discussed. The authors believe this contribution to be of crucial value for experts in the field as well as for policy makers deciding on the reimbursement of this powerful imaging modality.

A framework for voxel-based assessment of biological effect after proton radiotherapy in pediatric brain cancer patients using multi-modal imaging

INTRODUCTION

The exact dependence of biological effect on dose and linear energy transfer (LET) in human tissue when delivering proton therapy is unknown. In this study, we propose a framework for measuring this dependency using multi-modal image-based assays with deformable registrations within imaging sessions and across time.

MATERIALS AND METHODS

3T MRI scans were prospectively collected from 6 pediatric brain cancer patients before they underwent proton therapy treatment, and every 3 months for a year after treatment. Scans included T1-weighted with contrast enhancement (T1), T2-FLAIR (T2) and fractional anisotropy (FA) images. In addition, the planning CT, dose distributions and Monte Carlo-calculated LET distributions were collected. A multi-modal deformable image registration framework was used to create a dataset of dose, LET and imaging intensities at baseline and follow-up on a voxel-by-voxel basis. We modelled the biological effect of dose and LET from proton therapy using imaging changes over time as a surrogate for biological effect. We investigated various models to show the feasibility of the framework to model imaging changes. To account for interpatient and intrapatient variations, we used a nested generalized linear mixed regression model. The models were applied to predict imaging changes over time as a function of dose and LET for each modality.

RESULTS

Using the nested models to predict imaging changes, we saw a decrease in the FA signal as a function of dose; however, the signal increased with increasing LET. Similarly, we saw an increase in T2 signal as a function of dose, but a decrease in signal with LET. We saw no changes in T1 voxel values as a function of either dose or LET.

CONCLUSIONS

The imaging changes could successfully model biological effect as a function of dose and LET using our proposed framework. Due to the low number of patients, the imaging changes observed for FA and T2 scans were not marked enough to draw any firm conclusions.

Effect of uncertainties in quantitative 18 F-FDG PET/CT imaging feedback for intratumoral dose-response assessment and dose painting by number

PURPOSE

Intratumoral dose response can be detected using serial fluoro-2-deoxyglucose-(FDG) positron emission tomography (PET)/computed tomography (CT) imaging feedback during treatment and used to guide adaptive dose painting by number (DPbN). However, to reliably implement this technique, the effect of uncertainties in quantitative PET/CT imaging feedback on tumor voxel dose-response assessment and DPbN needs to be determined and reduced.

METHODS

Three major uncertainties, induced by (a) PET imaging partial volume effect (PVE) and (b) tumor deformable image registration (DIR), and (c) variation of the time interval between FDG injection and PET image acquisition (TI), were determined using serial FDG-PET/CT images acquired during chemoradiotherapy of 18 head and neck cancer patients. PET imaging PVE was simulated using the discrepancy between with and without iterative deconvolution-based PVE corrections. Effect of tumor DIR uncertainty was simulated using the discrepancy between two DIR algorithms, including one with and one without soft-tissue mechanical correction for the voxel displacement. The effect of TI variation was simulated using linear interpolation on the dual-point PET/CT images. Tumor voxel pretreatment metabolic activity (SUV₀ ) and dose-response matrix (DRM) discrepancies induced by each of the three uncertainties were quantified, respectively. Adverse effects of tumor voxel SUV₀ and DRM discrepancies on tumor control probability (TCP) in DPbN were assessed.

RESULTS

Partial volume effect and TI variations of 10 mins induced a mean ± standard deviation (SD) of tumor voxel SUV₀ discrepancies to be -0.7% ± 9.2% and 0% ± 4.8%, respectively. Tumor voxel DRM discrepancies induced by PVE, tumor DIR discrepancy, and TI variations were 0.6% ± 8.9%, 1.7% ± 9.1%, and 0% ± 7%, respectively. Partial volume effect induced SUV₀ and DRM discrepancies correlated significantly with the tumor shape and FDG uptake heterogeneity. Tumor DIR uncertainty-induced DRM discrepancy correlated significantly with the tumor volume and shrinkage during treatment. Among the three uncertainties, PVE dominated the adverse effects on the TCP, with a mean ± SD of TCP reduction to be 12.7% ± 9.8% for all tumors if no compensation was applied for.

CONCLUSIONS

Effect of uncertainties in quantitative FDG-PET/CT imaging feedback on intratumoral dose-response quantification was not negligible. These uncertainties primarily caused by PVE and tumor DIR were highly dependent on individual tumor shape, volume, shrinkage during treatment, and pretreatment SUV heterogeneity, which can be managed individually. The adverse effects of these uncertainties could be minimized by using proper PVE corrections and DIR methods and compensated for in the clinical implementation of DPbN.

Sensitivity analysis of FDG PET tumor voxel cluster radiomics and dosimetry for predicting mid-chemoradiation regional response of locally advanced lung cancer

We investigated the sensitivity of regional tumor response prediction to variability in voxel clustering techniques, imaging features, and machine learning algorithms in 25 patients with locally advanced non-small cell lung cancer (LA-NSCLC) enrolled on the FLARE-RT clinical trial. Metabolic tumor volumes (MTV) from pre-chemoradiation (PETpre) and mid-chemoradiation fluorodeoxyglucose-positron emission tomography (FDG PET) images (PETmid) were subdivided into K-means or hierarchical voxel clusters by standardized uptake values (SUV) and 3D-positions. MTV cluster separability was evaluated by CH index, and morphologic changes were captured by Dice similarity and centroid Euclidean distance. PETpre conventional features included SUVmean, MTV/MTV cluster size, and mean radiation dose. PETpre radiomics consisted of 41 intensity histogram and 3D texture features (PET Oncology Radiomics Test Suite) extracted from MTV or MTV clusters. Machine learning models (multiple linear regression, support vector regression, logistic regression, support vector machines) of conventional features or radiomic features were constructed to predict PETmid response. Leave-one-out-cross-validated root-mean-squared-error (RMSE) for continuous response regression (ΔSUVmean) and area-under-receiver-operating-characteristic-curve (AUC) for binary response classification were calculated. K-means MTV 2-clusters (MTVhi, MTVlo) achieved maximum CH index separability (Friedman p < 0.001). Between PETpre and PETmid, MTV cluster pairs overlapped (Dice 0.70-0.87) and migrated 0.6-1.1 cm. PETmid ΔSUVmean response prediction was superior in MTV and MTVlo (RMSE = 0.17-0.21) compared to MTVhi (RMSE = 0.42-0.52, Friedman p < 0.001). PETmid ΔSUVmean response class prediction performance trended higher in MTVlo (AUC = 0.83-0.88) compared to MTVhi (AUC = 0.44-0.58, Friedman p = 0.052). Models were more sensitive to MTV/MTV cluster regions (Friedman p = 0.026) than feature sets/algorithms (Wilcoxon signed-rank p = 0.36). Top-ranked radiomic features included GLZSM-LZHGE (large-zone-high-SUV), GTSDM-CP (cluster-prominence), GTSDM-CS (cluster-shade) and NGTDM-CNT (contrast). Top-ranked features were consistent between MTVhi and MTVlo cluster pairs but varied between MTVhi-MTVlo clusters, reflecting distinct regional radiomic phenotypes. Variability in tumor voxel cluster response prediction can inform robust radiomic target definition for risk-adaptive chemoradiation in patients with LA-NSCLC. FLARE-RT trial: NCT02773238.

Inter/intra-tumoral dose response variations assessed using FDG-PET/CT feedback images: Impact on tumor control and treatment dose prescription

PURPOSE

To quantify inter/intra-tumoral variations of baseline metabolic activity and dose response. To evaluate their impact on tumor control and treatment dose prescription strategies.

METHODS AND MATERIALS

Tumor voxel baseline metabolic activity, SUV₀, and dose response matrix, DRM, quantified using the pre-treatment and weekly FDG-PET/CT imaging feedback for each of 34 HNSCC patients (25 HPV+ and 9 HVP-) were evaluated. Inter/intra-tumoral variations of tumor voxel (SUV₀, DRM) for each of the HPV- and HPV+ tumor groups were quantified and used to evaluate the variations of individual tumor control probabilities and the efficiency of uniform vs non-uniform treatment dose prescription strategies.

RESULTS

Tumor voxel dose response variation of all tumor voxels assessed using FDG-PET/CT imaging feedback had the mean(CV) = 0.47(47%), which was consistent with those of previously published in vitro tumor clonogenic assay. The HPV- tumors had the mean(CV) dose response, 0.53(49%), significantly larger than those of the HPV+ tumors, 0.45(43%). However, their baseline SUVs were opposite, 6.5(56%) vs 7.7(65%). Comparing to the inter-tumoral variations, both HPV-/+ tumor groups showed larger intra-tumoral variations, (53%, 58%) vs (20%, 31%) for the baseline SUV and (38%, 37%) vs (31%, 21%) for the dose response. Due to the large dose response variations, treatment dose to control the tumor voxels has very broad range with CV of TCD₅₀ = 97% for the HPV- and 67% for the HPV+ tumor group respectively. As a consequence, heterogeneous prescription dose could potentially reduce the treatment integral dose for 92% of the HPV+ tumors and 78% of the HPV- tumors.

CONCLUSIONS

The study demonstrates that tumor dose response assessed using FDG-PET/CT feedback images had a similar distribution to those assessed conventionally using in vitro tumor clonogenic assay. Inter-tumoral dose response variation seems larger for HPV- tumors, but intra-tumoral dose response variations are similar for both HPV groups. These variations cause very large variation on the individual tumor control probability and limit the efficacy of dose escalation and de-escalation in conventional clinical practice. On the other hand, heterogeneous dose prescription guided by metabolic imaging feedback has a potential advantage in radiotherapy.