A deep learning framework for radiotherapy dose prediction: from MLC motion to patient-specific dose perturbation

The precision of modern radiotherapy hinges on the dynamic sculpting of dose by multi leaf collimators (MLCs). This study introduces a physics informed, interpretable framework that models MLC related delivery uncertainties and predicts their dosimetric impact, shifting patient-specific quality assurance (PSQA) from passive verification to proactive intervention. Approach: We analyzed trajectory log files from 1954 volumetric modulated arc therapy (VMAT) plans deliveries across two widely used MLC designs (Varian Clinac IX and Edge). Control point level MLC and gantry deviations were modeled using motion metrics (leaf speed, speed variation) and a novel temporal latency parameter (ΔtMLC) to capture system inherent uncertainties. These deviations were projected into the patient's 3D anatomy to create an MLC Position Deviation Texture (MPDT) to capture the spatial information of the deviation caused by MLC. MPDT and planning dose served as input to a dual pathway U-Net for predicting deliverable 3D dose distributions. The model was technically trained and tested on 1250 IX/200 Edge plans and clinically tested on an independent cohort of 428 IX/76 Edge cases under uncertainty aware conditions by structural similarity index measure (SSIM), percentage mean absolute error (PMAE) and deviation of gamma passing rate (GPR). Main results: MLC deviation patterns were design specific and exhibited consistent linear or quadratic relationships (average R² = 0.989). The MPDT enhanced framework achieved superior dose prediction accuracy (SSIM = 0.996, PMAE = 0.39%) compared to conventional planning dose only models (SSIM = 0.932, PMAE = 2.13%), along with improved GPR consistency (|ΔGPR| = 2.40% vs. 3.92%). In clinical testing, this performance advantage persisted. Significance: By integrating interpretable mechanical modeling with deep learning, this work establishes a scalable paradigm for PSQA in the era of precision oncology-one that traces dose deviations to their physical origins, quantifies delivery uncertainty, and enables proactive intervention.