Click here to


Are you sure ?

Yes, do it No, cancel

Voxel Forecast Classifier to Predict Spatially Variant Binary Tumor Voxel Response On Longitudinal FDG-PET/CT Imaging of FLARE-RT Protocol Patients

S Bowen1*, D Hippe1, W Chaovalitwongse2, P Thammasorn2, X Liu2, R Iranzad2, R Miyaoka1, H Vesselle1, P Kinahan1, R Rengan1, J Zeng1, (1) University of Washington, School of Medicine, Seattle, WA, (2) University of Arkansas, Fayetteville, AR


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Our Voxel Forecast algorithm was shown to predict spatially variant cancer therapy responses from multiscale clinical and imaging features with greater interpretability than deep learning approaches, but is restricted to continuous outcome variables. We extended the algorithm to predict binary tumor voxel response states (Voxel Forecast Classifier) during chemoradiation on longitudinal FDG-PET/CT imaging.

Methods: Twenty-five patients with locally advanced non-small cell lung cancer underwent baseline FDG-PET/CT imaging (PETpre) and during week 3 (PETmid) of chemoradiotherapy under the FLARE-RT trial (NCT02773238). PETpre, PETmid, and RTdose voxel grids were aligned and resampled within metabolic tumor volumes (MTV) to enable voxel tracking. MTV voxel response on PETmid was dichotomized by 30% SUV decrease relative to PETpre, analogous to PERCIST. Voxel Forecast variogram-weighted generalized linear regression was adapted for binary voxel response prediction using: (i) the logit link function between predictors and response, (ii) the binomial variance function in underlying equations to standardize residuals, (iii) the Newton-Raphson numerical solver. Patient- and voxel-level feature associations to binary voxel response were estimated by univariate and multivariate jack-knife odds ratios (OR). Prediction model performance was quantified by the area-under-ROC-curve (AUC) following leave-one-patient-out cross-validation (CV).

Results: Over 25 patients and 257k MTV voxels, 56% of voxels responded, with large inter-patient variability (standard deviation: 37%). Predictors of voxel response included PETpre MTV SUVmean (OR 2.2 [1.3-3.7], p=0.005), time from PETpre to PETmid (OR 1.7 [1.0-2.8], p=0.048), and PETpre voxel SUV (OR 1.4 [1.0-2.1], p=0.047). A parsimonious 3-feature Voxel Forecast Classifier model achieved CV-AUC of 0.72, compared to a univariate model with CV-AUC of 0.68.

Conclusions: Voxel Forecast Classifier was successfully implemented as an extension of Voxel Forecast to predict binary response status at patient and voxel scales. Voxel Forecast Classifier can decode spatial response and disease recurrence patterns to inform risk-adaptive radiotherapy in the setting of precision oncology.

Funding Support, Disclosures, and Conflict of Interest: Research supported by NIH/NCI R01CA204301.


PET, Image-guided Therapy, Quantitative Imaging


IM/TH- Image Analysis (Single Modality or Multi-Modality): Computer-aided decision support systems (detection, diagnosis, risk prediction, staging, treatment response assessment/monitoring, prognosis prediction)

Contact Email