Gamma Knife treatment planning using knowledge-based reinforcement learning

Clinical Snapshot

Key Findings

• P Value: Metastases: p=0.0136 (statistically significant). Acoustic neuromas: p=0.4493 (not statistically significant)

• Effect Size: Metastases: DRL agent mean score 3.925 ± 0.130 vs default 3.874 ± 0.147 (absolute difference +0.051). Acoustic neuromas: DRL agent mean score 4.035 ± 0.177 vs default 3.995 ± 0.365 (absolute difference +0.040). Score scale and clinical thresholds not defined in abstract.

• Primary Outcome: Composite plan quality score (derived from four dosimetric metrics: target coverage, selectivity, gradient index, and OAR dose) comparing DRL agent plans to default inverse planner slider plans, benchmarked against historical clinical plans

• Nnt Or Sensitivity: Not applicable in conventional sense; no NNT, sensitivity/specificity, or hazard ratio reported. No clinical outcome data available to calculate number needed to treat or similar metrics.

• Confidence Interval: Not reported for any outcome

Clinical Application

The DRL agent adjusts two priority weight sliders within the existing clinical inverse planner workflow, suggesting low integration burden if embedded within the Leksell planning software. However, clinical deployment would require prospective validation, regulatory approval, and institutional retraining of the model on site-specific historical plans. Planning time savings are not quantified. Clinician oversight protocols for agent-generated plans are not described. Gamma Knife radiosurgery is available at a small number of Australian tertiary centres including St Vincent's Hospital Sydney, Royal Melbourne Hospital, and Princess Alexandra Hospital Brisbane. The Leksell Gamma Knife Icon is the current generation platform used in Australia. TGA approval would be required before clinical deployment of any embedded planning software modification. RANZCR (Royal Australian and New Zealand College of Radiologists) and ACPSEM (Australasian College of Physical Scientists and Engineers in Medicine) guidelines on stereotactic radiosurgery quality assurance would apply. PBS does not directly fund radiosurgery planning software; costs would be absorbed by hospital or state health system budgets. Australian centres considering adoption would need to retrain the model on their own institutional planning data, as the reward function is site-specific. Given the small number of Australian Gamma Knife centres, multicentre data pooling would be necessary to achieve adequate training set sizes, particularly for acoustic neuromas. Patients with single brain metastases or acoustic neuromas undergoing Gamma Knife stereotactic radiosurgery at centres using the Leksell Gamma Knife Icon platform with its proprietary inverse planner. Not yet validated for multi-target metastases, arteriovenous malformations, trigeminal neuralgia, or other Gamma Knife indications.

Abstract

BACKGROUND: Inverse planning is often used for Gamma Knife radiosurgery, allowing clinicians to mathematically specify desired clinical objectives and dose limits. The objectives are controlled by weights that are manually tuned to find the desired trade-off, which varies from case to case. Automation of this process can reduce clinical workload and improve consistency in plan quality. PURPOSE: To train a deep reinforcement learning agent using a reward function that incorporates the clinical metrics from past plans into its scoring criteria. The metric trade-off from the clinical plan is scored higher than all others, guiding the agent to produce plans with similar trade-offs. METHODS: An agent was trained to adjust the two priority weights (i.e., digital slider bars) in the clinical inverse planner. The agent consists of a neural network that receives the metrics and dose distribution of the current plan and the target and organ-at-risk masks as inputs. These methods were demonstrated on a dataset of 204 single-target metastases and a dataset of 71 acoustic neuroma cases. The cases were split into training, validation, and testing sets of size 123/41/40 and 42/14/15 for the metastases and acoustic neuromas, respectively. RESULTS: On the metastases test dataset, the agent achieved a significantly higher (p = 0.0136) average plan score (3.925 ± 0.130) compared to the default slider plans (3.874 ± 0.147). On the acoustic neuromas test dataset, the agent achieved a higher (p = 0.4493) average plan score (4.035 ± 0.177) compared to the default slider plans (3.995 ± 0.365). The higher plan scores are reflected in the four plan quality metrics: the agent's plans, on average, had metrics more similar to the clinical plans, compared to the default slider plans, for both test datasets. CONCLUSIONS: The proposed reward function enabled the agent to learn to find plans that aligned with historical planning decisions. Future work will investigate providing the agent with additional inputs that can explain the variability in planning decisions, which would further improve its performance.

References

1. 1.Huynh, C., Ahlgren, B., Zhang, B., Fay, D., Nordström, H., & Ruschin, M. (2026). Gamma Knife treatment planning using knowledge-based reinforcement learning. Medical Physics. https://doi.org/10.1109/TRPMS.2021.3107454