335 - Predicting Voxel-Level Local Progression for Locally Advanced NSCLC following Chemoradiation Using Multi-Task Deep Learning

Purpose/Objective(s): Current patient-level predictions of local progression in non-small cell lung cancer (NSCLC) lack the spatial resolution to guide personalized treatment. This study aims to predict voxel-level local progression (vox-LP) following chemoradiation using 3D imaging and dose inputs, enabling the identification of high-risk LP regions for precise spatial dose escalation and guiding LP monitoring during follow-up for NSCLC.

Materials/Methods: Data from 46 NSCLC patients in the FLARE-RT trial (NCT02773238) were analyzed, where biologically guided dose escalation to tumors provided unique variability for vox-LP prediction. After excluding 13 patients due to early censoring, 33 patients (8 with LP) were included in the study. Vox-LP was defined radiographically via serial CT and PET imaging, with recurrence contours delineated by board-certified radiation oncologists. Pre-treatment [¹8F] FDG-PET scans, planning CT, and dose data were fed into a composite convolutional neural network (CNN) with adaptive-weighted multi-task learning that was benchmarked against baseline logistic regression (LR). The composite CNN simultaneously optimized three interrelated tasks: vox-LP risk prediction, patient-level LP classification, and patient-level LP volume percentage regression. A 5-fold stratified cross-validation was performed. Vox-LP prediction performance was assessed using cross-validated AUC across all patients, as well as patient-specific AUC and Dice coefficients (DSC) for all positive cases.

Results: Our proposed CNN outperformed LR across all patients (cross-validated AUC: 0.793 vs. 0.748, p<0.001) and in LP-positive patients (mean patient-specific AUC: 0.774 vs. 0.736, p<0.001; mean DSC: 0.300 vs. 0.259). Specifically, it achieved significantly higher AUCs in 7 LP-positive cases and equivalent AUC in the remaining LP-positive case. Also, it achieved higher DSCs in 6 LP-positive cases. Interestingly, in case #4 where LR outperformed the CNN by DSC, both underperformed overall (AUC < 0.5, DSC < 0.01), suggesting that neither model was effective. Overall, CNN demonstrated better discrimination and spatial overlap in vox-LP prediction compared to LR.

Conclusion: This study demonstrates the feasibility of vox-LP prediction for NSCLC using multitask CNNs, representing a significant advancement over traditional statistical models. By identifying high-risk subregions prior to treatment, this approach could enable personalized radiation dose sculpting to enhance local control and guide monitoring of these regions on follow-up imaging.