Technical Note: A cascade 3D U-Net for dose prediction in radiotherapy.

Area of Science:

• Medical Physics
• Artificial Intelligence in Healthcare
• Radiotherapy Planning

Background:

• Accurate three-dimensional (3D) dose prediction is crucial for effective radiotherapy.
• Learning robust dose prediction models from limited datasets presents a significant challenge in radiation therapy.

Purpose of the Study:

• To develop a precise 3D dose prediction model using cascaded deep learning and advanced training strategies with limited data.
• To overcome the limitations of small datasets in achieving accurate radiotherapy dose predictions.

Main Methods:

• A Cascade 3D (C3D) model was developed, integrating cascade mechanisms with 3D U-Net architectures.
• Data augmentation and knowledge distillation techniques were employed to enhance model generalization and learning capabilities.
• The C3D model was evaluated on the OpenKBP challenge dataset and compared against leading dose prediction models.

Main Results:

• The C3D model achieved a mean absolute error (MAE) of 2.50 Gy in nonzero dose areas, outperforming five other models.
• The C3D model secured first place in both dose and DVH (Dose-Volume Histogram) streams of the AAPM 2020 OpenKBP challenge.
• Performance was evaluated using voxel-based MAE and clinical dosimetric metrics.

Conclusions:

• Cascading U-Nets offer an effective solution for 3D dose prediction, particularly with limited datasets.
• Data preprocessing, augmentation, and optimization procedures are more critical than deep learning network architectural modifications for robust dose prediction.