Author: Stephen R. Bowen, Chunyan Duan, Daniel S. Hippe, Qiantuo Liu, Jing Sun, Jiajie Wang, Shouyi Wang, Faisal Yaseen, Xiaojing Zhu 👨🔬
Affiliation: Tongji University, University of Washington, Department of Radiation Oncology, Fred Hutchinson Cancer Center, University of Washington, Shanghai University of Electric Power, Fred Hutchinson Cancer Center, University of Texas at Arlington 🌍
Purpose: Accurate prediction of patient response to radiotherapy plays a crucial role in monitoring disease progression and assessing treatment efficacy, enabling clinicians to develop personalized therapeutic strategies. We constructed a multi-task learning model for mid-treatment FDG PET response prediction.
Methods: We developed a two-layer, two-task prediction model for tumour radiotherapy response based on 3D imaging and residual networks (3D ResMMoE) for mid-treatment FDG PET response prediction in 23 patients (Clinical trial: NCT02773238) with locally advanced non-small cell lung cancer (LA-NSCLC). 3D ResMMoE model consists of two main parts: (i) a hard parameter sharing layer, which uses a three-dimensional convolutional neural network (3D CNN) to extract features from fused pre-treatment PET images (PETpre) and radiotherapy dose (Dose) images; (ii) a soft parameter sharing layer, which inputs the extracted features into a Multi-gate Mixture-of-Experts (MMoE) model, and simultaneously predicts the mean standard uptake value (SUV) on mid-treatment images (PETmid) and the change in SUV between PETmid and PETpre ([SUVmean_mid–SUVmean_pre]/ SUVmean_pre). The 3D ResMMoE model is benchmarked against classical deep learning models and intermediate models (3D MMoE, 3D CNN, 3D CNN with hard sharing mechanism), with performance quantified by root mean square error (RMSE) following leave-one-patient-out cross-validation.
Results: In predicting the mean SUV on PETmid, the RMSE of the 3D ResMMoE model (1.60 SUV) was numerically lower than the RMSE from each benchmark model (RMSE: 1.65-1.73 SUV). In predicting the change in mean SUV between PETmid and PETpre, the RMSE of the 3D ResMMoE model (0.20) was numerically better than the RMSE from each benchmark model (RMSE: 0.22-0.26).
Conclusion: We propose the 3D ResMMoE model to achieve the dual-task regression prediction of the mean SUV on PETmid and the change in mean SUV between PETmid and PETpre, which improves the accuracy and stability of radiotherapy response prediction to adapt treatment plans.