Author: Eric Aliotta, Michalis Aristophanous, Joseph O. Deasy, Bill Diplas, Milan Grkovski, James Han, Vaios Hatzoglou, Jeho Jeong, Nancy Y Lee, Ramesh Paudyal, Nadeem Riaz, Heiko Schoder, Amita Shukla-Dave 👨🔬
Affiliation: Department of Radiology, Memorial Sloan Kettering Cancer Center, Department of Radiation Oncology, Memorial Sloan Kettering Cancer Center, Department of Medical Physics, Memorial Sloan Kettering Cancer Center 🌍
Purpose: To forecast radiotherapy treatment response for head and neck cancer (HNC) using multimodality imaging and personalized radiobiological modeling.
Methods: Multi-modality imaging data from diffusion weighted-magnetic resonance imaging (DW-MRI) and positron emission tomography (PET) with Fluorodeoxyglucose (FDG) and Fluoromisonidazole (FMISO) tracers were collected from N=72 patients undergoing chemo-radiotherapy for human papilloma virus associated HNC. A personalized treatment response prediction framework was then developed using a mechanistic tumor control probability model. Model input parameters (growth fraction – GF, cell loss factor – CLF, and cell-cycle time – Tc) were initialized using pre-treatment imaging metrics (DW-MRI Dmean, FDG SUVmean, FMISO SUVmean) within nodal gross tumor volumes (GTV) and updated to reflect observed longitudinal intra-treatment changes in GTV volume (defined on weekly MRI). Prediction accuracy was assessed through five-fold cross-validation using mean absolute error (MAE) of weekly volume predictions, linear correlation with measured volumes, and locoregional recurrence (LRR) prediction.
Results: Personalized modeling based on pre-treatment imaging significantly improved longitudinal volume prediction accuracy and correlation with measurement compared with a generic population model (MAE=23.4±10.0% vs. 24.9±9.0%, p=0.002 on paired t-test, R=0.82 vs. 0.72). When incorporating feedback from longitudinal measurements, penalizing large deviations from pre-treatment model parameters using variational regularization was necessary to maintain stability and models tuned using week-1 and week-2 volume measurements further improved accuracy in subsequent weeks (MAE=12.5±8.1%, 10.7±9.9%, R=0.91, 0.97). Model-predicted volumes based on baseline+week-1 information significantly improved LRR prediction compared with week-1 volume data alone (area under the curve, AUC=0.83 vs. 0.77, p=0.03) and was similar to prediction using week-3 volume data alone (AUC=0.83 vs. 0.85, p=non-significant).
Conclusion: Integrating multimodality quantitative imaging into radiobiological digital twin models improved treatment response prediction for HNC patients undergoing chemo-radiotherapy. Personalized models improved predictions of intra-treatment regression compared with a population model and further improvements were observed by refining models with early-treatment volume measurements.