Impact of Auxiliary Task Heterogeneity on Dose Prediction: A Controlled Study of Optimization Coupling and Dose Outcomes.

Authors

Abstract

Auxiliary-task training has been increasingly explored in radiotherapy dose prediction to incorporate planning-relevant priors, such as patient anatomy, beam geometry, and spatial information, into model optimization. However, when auxiliary tasks differ in their physical meaning, input representation, and prediction target, their roles cannot be adequately understood from final prediction performance alone. Particularly in auxiliary-supervision frameworks without shared feature representations, it remains unclear whether an auxiliary task establishes an effective optimization connection with the primary network, actually uses the additional information provided to it, and produces stable performance effects. The purpose of this study was to compare heterogeneous auxiliary tasks for left-sided breast IMRT dose prediction under controlled conditions and to evaluate their relationships with the primary dose-prediction task in terms of optimization coupling, information use, and performance robustness. We implemented an independently parameterized auxiliary-supervised joint optimization framework comprising a primary 3D dose prediction network and three auxiliary networks targeting patient anatomy, beam geometry, and isocenter-related spatial information, without shared encoders or intermediate features. Specifically, we evaluated a dose-only baseline (DS) and four auxiliary-supervision configurations combining dose prediction with beam-geometry (BS), center-related (CS), patient-structure (PS), or all auxiliary supervision (AS). Using data from 30 patients, we generated 130 plans per patient via bounded perturbations of beam angles and isocenter position, and evaluated all configurations on 650 test plans from held-out patients under identical training settings. Prediction performance was first assessed using voxel-wise errors and dosimetric endpoints. Gradient analysis was then used to quantify direct optimization coupling between auxiliary losses and DoseNet, while a matched CT + predicted-dose versus CT-only input ablation evaluated the use of predicted-dose information by the Beam and Center tasks. Finally, performance robustness was assessed across multiple random seeds and under composite-loss and MSE-only primary dose objectives. DS achieved the lowest overall MAE/RMSE, while BS showed comparable performance: under the composite-loss setting, BS exhibited more consistent cross-seed improvements across several clinically relevant dosimetric endpoints. Further analyses revealed distinct pathways among the auxiliary tasks. Both Beam and Center were directly connected to DoseNet through the predicted dose, but Beam showed substantially larger relative gradient magnitudes, and inclusion of the predicted dose improved Beam prediction across all three random seeds. In contrast, Center showed no consistent benefit from predicted-dose information, while Patient had no direct auxiliary-gradient pathway to DoseNet. Changing the primary dose objective from the composite loss to MSE-only substantially altered the endpoint-level effects and cross-seed consistency of the auxiliary configurations, with the most pronounced loss of consistency observed for BS. The role of auxiliary supervision cannot be determined from final performance changes alone. Optimization coupling, information use, and performance robustness represent distinct dimensions that should be evaluated separately. Gradient analysis, input ablation, performance comparisons across random seeds and primary loss formulations together provide a more complete framework for evaluating auxiliary supervision than performance comparison alone.

Tags

Topics