Dual Path Networks for Voxel-Level Prediction of Mid-Chemoradiotherapy Response In FDG PET for Locally Advanced Non-Small Cell Lung Cancer
Abstract
Purpose
Accurate prediction of tumor response during chemoradiotherapy is essential for treatment optimization but remains challenging. We developed a deep learning model based on a Dual Path Network (DPN), which is a hybrid architecture combining elements of ResNet and DenseNet and designed with a voxel-level output structure to predict tumor response on FDG PET imaging during treatment.
Methods
25 locally advanced non-small cell lung cancer patients undergoing chemoradiotherapy on the clinical trial (NCT02773238) were included in the analysis. We employed a deep learning architecture based on DPN to predict voxel-wise mid-treatment SUV (MidSUV) and voxel-wise mid-treatment SUV response ratio, calculated from the MidSUV and pre-treatment SUV (PreSUV) as ΔSUV (ΔSUV = (PreSUV - MidSUV)/PreSUV). We implemented a leave-one-out cross-validation strategy across 25 patients, resulting in 25 rounds of experiments. In each round, one patient was sequentially selected as the test patient. From the remaining cohort, one patient from each of the three response categories (classified according to SUV changes: high [n=8], medium [n=8], and low [n=9]) was randomly selected to form the validation set, while the rest comprised the training set. Prediction accuracy was quantified using the root mean squared error (RMSE) and the mean absolute error (MAE) between predicted and reference values, and Spearman's rank correlation coefficient was used to assess monotonic association; all metrics were computed for both voxel-wise MidSUV and ΔSUV predictions.
Results
Evaluating the best-performing model from each validation round, the RMSE was 1.26 for MidSUV and 0.26 for ΔSUV, while the MAE was 0.99 and 0.19. Spearman’s rank correlation coefficient was 0.69 for MidSUV and 0.15 for ΔSUV across all patients.
Conclusion
In this study, we employed DPN to predict lung tumor response following chemoradiotherapy. By utilizing an innovative network architecture, we achieved voxel-level prediction capability, which can guide future patient-level treatment strategies.