Treatment plan prediction for lung IMRT using deep learning based fluence map generation.

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Automated treatment planning using convolutional neural networks (CNNs) shows promise for direct fluence prediction.
• Previous studies often used fixed dose prescriptions and beam angles, limiting clinical applicability.
• Adapting these methods for complex indications like lung cancer requires addressing variable parameters.

Purpose of the Study:

• To develop and evaluate automated treatment planning for lung cancer using CNNs with variable dose prescriptions and beam angles.
• To investigate the impact of clinical collimator angles and various input parameters on fluence prediction accuracy.
• To assess the performance of a complete, user-independent planning workflow.

Main Methods:

• A dataset of 152 lung cancer patients treated with Intensity-Modulated Radiation Therapy (IMRT) was used.
• Two CNNs were compared: one using standard collimator angles and another using personalized, clinical angles.
• CNNs were trained with various combinations of CT and contour inputs, and a full user-free workflow was evaluated.

Main Results:

• Fluence prediction CNNs achieved comparable accuracy (mean absolute error) with both fixed and variable collimator angles, with negligible differences in DVH metrics.
• The impact of anatomical inputs on prediction accuracy was found to be minimal.
• A complete user-free planning workflow showed increased DVH differences, indicating areas for improvement.

Conclusions:

• Personalized collimator angles in CNN-based fluence prediction perform comparably to standard fixed angles for lung cancer treatment planning.
• Anatomical inputs have a limited effect on the accuracy of fluence prediction.
• Combining dose and fluence prediction CNNs negatively impacted fluence prediction accuracy, necessitating further investigation.