A Preliminary Study on Deep Learning-Based Plan Quality Prediction in Gamma Knife Radiosurgery for Brain Metastases

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence

Background:

• Gamma Knife (GK) plan quality is influenced by lesion characteristics, making direct metric comparison difficult across patients.
• Patient-specific geometry significantly impacts treatment outcomes and plan evaluation.

Purpose of the Study:

• To develop a deep learning-based method for predicting achievable, clinically acceptable GK plan quality from patient geometry.
• To enhance the accuracy and robustness of plan quality metric predictions for brain metastases.

Main Methods:

• A hierarchically densely connected U-Net (HD-U-Net) was trained to predict 3D dose distributions and plan quality metrics (coverage, selectivity, GI, CI50).
• Incorporated Dice similarity coefficient losses with Mean Squared Error (MSE) loss to improve prediction accuracy.
• Validated using ten-fold cross-validation on 463 brain metastases (BMs) from 175 patients.

Main Results:

• The proposed method demonstrated smaller mean absolute errors across all evaluated plan quality metrics compared to the baseline HD-U-Net.
• Significant improvements were observed for small metastases, and primarily in Gradient Index (GI) and Conformity Index at 50% prescription dose (CI50) for larger lesions.
• Statistical significance (p < 0.05) was confirmed using paired Wilcoxon signed-rank tests.

Conclusions:

• The deep learning approach outperforms the baseline in predicting overall trends and per-lesion accuracy for GK plan quality.
• The method offers potential as a pre-planning tool for constraint setting and a post-planning tool for quality control.
• This work provides a foundation for developing automated planning systems using deep reinforcement learning for GK treatments.