Problem

Conventional radiation therapy (RT) planning is a time-consuming and iterative process that heavily relies on expert planners. Existing automated AI-based methods, including reinforcement learning (RL) and model predictive control (MPC), often still require multiple dose evaluations and interactions with the treatment planning system (TPS), leading to planning times of several minutes. This limits the widespread availability of fast, consistent, and high-quality treatment planning, especially in regions with a scarcity of qualified personnel. The challenge is to develop an AI method that can generate high-quality, clinically deliverable RT plans much faster, without iterative TPS interactions.

Method

The authors introduce AIRT (Artificial Intelligence-based Radiotherapy), an end-to-end deep-learning framework designed to directly infer deliverable single-arc VMAT prostate treatment plans from CT images and structure contours. The core of AIRT is a feed-forward pipeline comprising several modules: a Dose Proposer (3D ResUNet) to predict an initial dose distribution, a Beam’s Eye View (BEV) Projection module, a Bev2Fluence network (MedNeXT backbone) to predict initial fluence maps, a differentiable deep-learning dose computation engine to evaluate the delivered dose, a dose error correction (Err) module, a second Bev2Fluence Correction network to refine fluence maps based on the error, and a rule-based leaf sequencer to convert fluence maps into MLC sequences. A key innovation is the differentiable dose feedback mechanism that refines fluence maps in a single pass to improve target homogeneity and OAR sparing. Additionally, an adversarial loss with a discriminator network is used during training to ensure the generated fluence maps are deliverable and maintain the characteristic two-level pattern of VMAT fluence. The model was trained in two stages on over 10,000 augmented intact prostate cases derived from a base dataset of 1,277 CT scans, using various reconstruction losses and finally incorporating the adversarial loss and OAR sparing controls.

Results

AIRT demonstrated the ability to generate single-arc VMAT prostate plans, including leaf sequencing, in under one second on a single Nvidia A100 GPU. When evaluated using the AcurosXB dose engine, AIRT plans showed similar target homogeneity (HI = 0.10 ± 0.01) and comparable OAR sparing metrics to clinical RapidPlan Eclipse plans, satisfying non-inferiority test margins (p < 0.05). An ablation study confirmed the importance of both the differentiable dose feedback and adversarial loss for achieving good PTV homogeneity and leaf sequenceability. The framework also showed flexibility in adapting to user-defined OAR sparing preferences at inference time without retraining. The most computationally intensive step within the pipeline was the deep-learning-based dose calculation, taking 654 ms.

Implications

This ultra-fast, end-to-end AI planning pipeline represents a significant advancement towards standardized and streamlined radiation treatment planning. Reducing planning times from minutes to under a second could revolutionize clinical workflows by increasing throughput, ensuring consistent planning quality derived from large datasets, and enabling interactive exploration of multiple candidate plans by clinicians in real-time. The ability to export DICOM RT Plans facilitates easy integration into existing clinical systems. Future work aims to generalize AIRT to other body regions, multi-arc VMAT, and more complex dose objectives. While the current dose engine is an approximation and operates at a slightly coarser resolution (4 mm), further alignment with clinical solvers like AcurosXB and improvements in differentiable dose engine efficiency could enhance plan quality and resolution. Independent testing has shown comparable target coverage to manually crafted plans and successful patient-specific quality assurance.