Patient-Specific Deep Learning–Guided Automatic Planning for Head and Neck Adaptive Radiotherapy
Abstract
Purpose
Deep learning dose prediction (DP) models rapidly predict radiotherapy dose distributions, addressing critical time constraints in adaptive radiotherapy (ART). Recent patient-specific (PS) approaches fine-tune models on individual initial plans for highly tailored predictions. However, the optimal PS approach remains unknown, and their clinical utility is yet to be established. This study compares PS model performance and demonstrates applications to automated planning in head-and-neck ART through isodose-based mimicking on the Ethos Intelligent Optimization Engine (IOE).
Methods
76 head-and-neck patients with repeated CT scans and treatment plans were included: 57 for training, 10 for validation, and 9 for testing. Four DP strategies were evaluated: (1) conventional training (CONV) on the entire training set; (2) PS intentional deep overfitting learning (IDOL), fine-tuning CONV on each patient's initial plan; (3) IDOL with frozen encoder (IDOL_FrozenEnc), preventing encoder weight updates during fine-tuning; and (4) single-patient learning (SingleP), trained only on test-patient initial plans. The best-performing model generated predictions for isodose-based mimicking on IOE to automatically create treatment plans, further compared against template-based and manually-optimized plans.
Results
IDOL_FrozenEnc and IDOL achieved the best DP performance, with median prediction errors below 2 Gy for mean dose and below 2.5Gy for Dmax metrics. SingleP showed substantially worse performance (8.1Gy spinal canal Dmax error, 12.3Gy mandible Dmax error), while CONV performed intermediately. IDOL_FrozenEnc was 0.5Gy lower PTV D95 error compared to IDOL justifying its selection for automatic planning. All manual replans met top priority objectives (target constraints and spinal cord/brainstem limits) across 9 test patients. Template-based plans failed 2 objectives, i.e., one above spinal cord maximum dose threshold and one above PTV upper dose threshold. IDOL_FrozenEnc-mimicked plans met all top priority objectives.
Conclusion
IDOL_FrozenEnc outperformed both CONV and other PS approaches. When combined with isodose-based mimicking, the generated plans surpassed template-based ones by meeting more clinical objectives.