Radiotherapy treatment planning is a critical and complex process that determines how radiation will be delivered to a patient's tumor while minimizing dose to surrounding healthy tissue. Conventional planning requires highly trained medical dosimetrists to iteratively adjust parameters in treatment planning software, a process that is simultaneously a science and an art, heavily dependent on individual expertise and experience.
Artificial intelligence has demonstrated transformative potential across numerous fields, and AI-based treatment planning is emerging as a key application in radiation oncology. Automated planning tools can spare dosimetrists from manual optimization tasks, improve plan consistency across patients, reduce planning time, and potentially enable individualized dose escalation strategies that would be too time-consuming to implement through manual methods alone.
Current AI approaches in radiotherapy treatment planning can be categorized into three main types: automated rule implementation and reasoning (ARIR), which encodes clinical guidelines as automated rules; knowledge-based planning (KBP), which learns from historical treatment plans to predict optimal dose distributions; and multicriteria optimization, which balances competing clinical objectives. Emerging approaches include voxel-based dose prediction using deep learning and reinforcement learning-based planning.
A rigorous comparison of two widely used commercial treatment planning automation algorithms was conducted: Pinnacle AutoPlan (an ARIR-based system) and Varian RapidPlan (a KBP-based system). The study was designed to minimize human factors and ensure objective comparison by using an established quantitative planning quality metric, addressing a critical gap in the literature where no well-controlled head-to-head comparison had previously been reported.
Both algorithms were evaluated on prostate bed planning scenarios including a case from a professional human planner competition. Notably, both automation algorithms achieved above-average human planner performance on the competition case, with marked efficiency improvements over manual planning. This finding demonstrates that automated planning has already reached or exceeded typical human expert performance on standardized planning tasks.
Despite testing two fundamentally different algorithmic approaches - one encoding clinical rules explicitly versus one learning patterns from prior plans - both systems yielded similar performance by quantitative metrics. This equivalence suggests that the underlying quality of clinical planning rules and the quality of historical training data may matter more than the specific algorithmic architecture for many standard treatment sites.
AI automation was applied to explore individualized dose escalation in pancreatic stereotactic body radiotherapy (SBRT). Pancreatic SBRT has shown promise for borderline resectable and locally advanced pancreatic cancer, but individualized target dose escalation within normal tissue dose limits has historically been too time-consuming to be practical in routine clinical workflows. Using the automated ARIR algorithm Pinnacle AutoPlan, the dose-escalation limit was systematically explored across patients, demonstrating that AI makes clinically meaningful personalized dose optimization feasible.
A complementary study automated and optimized beam settings for whole-breast radiotherapy, a task that typically requires a medical dosimetrist to spend substantial time on manual forward planning. The automated approach optimized beam angles and orientations followed by automatic fluence optimization, demonstrating that AI can manage even the upstream beam geometry decisions, not just the dose optimization stage of planning.
Collimator setting optimization for pancreatic SBRT with volumetric-modulated arc therapy (VMAT) was explored, targeting a parameter that is usually not optimized in conventional clinical practice. The optimization algorithm discovered settings that achieved significantly better sparing of organs at risk, illustrating that AI can find improvements in degrees of freedom that human planners may not systematically explore due to time constraints.
A systematic comparison of three KBP dose prediction algorithms on head and neck, lung, and prostate cases revealed an important finding about data requirements: when patient data are limited, simpler statistical learning methods are more robust to patient variability and outperform more sophisticated machine learning approaches for dose prediction. This counterintuitive result has important practical implications for smaller radiotherapy centers with limited historical plan databases.
To address robustness challenges with novel patient anatomies not well-represented in training data, a case-based reasoning approach was developed that combines atlas-based and regression-based dose prediction. This hybrid approach improves prediction robustness for patients whose anatomy falls outside the distribution of the training database, addressing a key limitation of purely data-driven KBP methods.
A knowledge-based statistical inference method using a large cohort of 927 head and neck cancer patients was used to evaluate treatment plan quality by selecting historically similar plans from the database based on both geometry and dosimetry. This work developed infrastructure for automatic data extraction, anonymization, and analysis, laying the groundwork for multi-institutional data integration to further expand training databases and improve prediction reliability.
Fully automated treatment planning requires automated delineation of regions of interest (ROIs) including tumor targets and organs at risk. Manual contouring is one of the most time-consuming steps in the planning workflow, often requiring several hours per patient. Conventional mathematical segmentation algorithms (intensity thresholding, edge detection) work well for anatomically distinct structures like lung and spinal cord but have limited accuracy for complex target volumes.
A deep learning U-Net model was applied to automatically segment tumor targets for nasopharyngeal radiotherapy from CT images. Trained on 502 patients, the model segmented tumor targets in under one minute, achieving over 70% agreement (Dice similarity coefficient) for primary tumors and over 60% for involved lymph nodes compared to manual segmentation that typically requires several hours. While encouraging, the remaining 30-40% inaccuracy required manual touch-up, ultimately saving less than 30 minutes per case compared to fully manual contouring.
A transfer learning approach was applied to a deep CNN for detecting malignant prostate lesions on multiparametric MRI, using a model pre-trained on ImageNet and fine-tuned for the clinical task. The best-performing model achieved the third highest score among 72 submitted models from 33 teams in the Prostate-X Grand Challenge, with performance comparable to radiologists following standard clinical protocol - demonstrating that transfer learning can effectively compensate for the relatively small size of labeled medical imaging datasets.
Data size is a fundamental challenge for AI in radiotherapy treatment planning. Treatment planning datasets are smaller than other healthcare AI applications such as medical imaging, creating uncertainty in data-driven models and limiting generalizability. Data heterogeneity within single-institution datasets - arising from variations in scanner settings, contouring styles, and treatment techniques over time - adds further complexity that must be addressed through careful data homogenization before model training.
Model selection based on available data is critical: the finding that simpler statistical methods outperform more complex machine learning under limited data directly challenges the assumption that more sophisticated algorithms are universally preferable. Clinical centers implementing AI planning tools must carefully match algorithm complexity to available data volume and quality, prioritizing robustness over raw performance metrics.
Realizing the vision of fully automated treatment planning requires integration of automated segmentation, automated planning, and automated plan quality evaluation into a seamless workflow. Current research demonstrates promising results across individual components, but combining them into a reliable end-to-end system that meets clinical quality standards across diverse patient presentations remains an active research challenge requiring ongoing collaboration between AI researchers, radiation physicists, and oncologists.
The primary clinical benefit of AI treatment planning is workflow efficiency: by automating time-consuming planning steps, dosimetrists and physicists can focus their expertise on cases requiring special attention, improving throughput while maintaining or improving plan quality. Studies consistently show that automated plans achieve performance matching or exceeding average human planner quality, with dramatically reduced planning time.
AI-enabled dose escalation and parameter optimization open opportunities for personalized treatment intensification that would be impractical through purely manual methods. For cancer sites like locally advanced pancreatic cancer where treatment outcomes remain poor, systematic individualized dose escalation enabled by AI could translate directly to improved tumor control and potentially patient survival - a compelling clinical application of treatment planning automation.
Future directions highlighted in this collection include deep learning for voxel-based dose prediction, which can generate entire dose distributions directly from patient anatomy without iterative optimization; reinforcement learning approaches that learn planning strategies through trial-and-error optimization; and multi-institutional data sharing frameworks that expand training datasets while maintaining patient privacy through anonymization infrastructure. Regulatory frameworks and standards for validating AI planning tools in clinical practice are also identified as essential requirements for broad adoption.
AI is rapidly advancing from supporting individual steps in the radiotherapy planning workflow toward enabling fully automated, high-quality treatment planning across cancer sites. The reviewed studies demonstrate that automated algorithms have reached or exceeded average human planner performance on standardized tasks, while enabling clinically meaningful applications like individualized dose escalation that were previously impractical.
Key insights from this collection include the importance of matching algorithm complexity to available data, the complementary roles of rule-based and knowledge-based approaches, and the transformative potential of deep learning for segmentation and dose prediction. Each finding contributes to a clearer roadmap for moving AI-based treatment planning from research to routine clinical practice.
Like self-driving cars reshaping transportation, AI-based radiotherapy planning is poised to fundamentally reshape the roles of human planners and physicists in the treatment planning process - not by replacing human expertise, but by augmenting it with unprecedented speed, consistency, and optimization capability, ultimately enabling more effective and personalized cancer treatment for more patients.