Accuracy Improvement Method Based on Characteristic Database Classification for IMRT Dose Prediction in Cervical...

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• Deep learning (DL) models require large, high-quality datasets for optimal performance, which are challenging to obtain for rare clinical scenarios.
• Effective training data collection strategies are essential for developing accurate DL-based dose prediction models in radiotherapy.

Purpose of the Study:

• To identify optimal training data collection strategies for deep learning (DL) based dose prediction models in intensity-modulated radiation therapy (IMRT).
• To evaluate the impact of dataset homogeneity, beam settings, and proposed methods on DL dose prediction accuracy.

Main Methods:

• Utilized 325 clinically approved cervical IMRT plans for comparative experiments.
• Investigated the impact of beam angles, number of beams, and patient position on DL dose prediction.
• Proposed a novel geometry-based beam mask generation method and a "full-database pre-trained strategy" for model training.

Main Results:

• Models trained on homogeneous datasets achieved the best performance (e.g., PTV error: 0.29 ± 0.15%).
• A homogeneous dataset is more accessible for training accurate dose prediction models compared to non-homogeneous ones.
• The proposed beam mask consistently improved model performance, especially for datasets with varying beam settings (reducing PTV error from 0.8 ± 0.14% to 0.29 ± 0.15%).

Conclusions:

• Recommends consistent datasets for patient-specific IMRT dose prediction models.
• Suggests that large datasets with diverse beam angles, coupled with beam information in training, can yield good models when consistency is not feasible.
• Highlights the effectiveness of the proposed beam mask and full-database pre-training strategies for improving DL dose prediction accuracy.