An ultra-fast deep-learning-based dose engine for prostate VMAT via knowledge distillation framework with limited...

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Traditional dose calculation algorithms in radiotherapy face accuracy-efficiency trade-offs.
• Current deep-learning (DL) dose engines are computationally intensive, limiting clinical use on resource-constrained devices.

Purpose of the Study:

• To develop a compact DL-based dose engine using knowledge distillation (KD) for ultra-fast and accurate dose calculation in prostate Volumetric Modulated Arc Therapy (VMAT).
• To reduce computational complexity and hardware requirements for broader clinical applicability.

Main Methods:

• A knowledge distillation (KD) framework was employed, training a smaller 'student' model using a larger 'teacher' model.
• The model input comprised patient CT scans and VMAT dose data; the output was a DL-calculated patient dose.
• Ground-truth doses were obtained using Monte Carlo simulations; model performance was assessed via Gamma analysis and inference efficiency.

Main Results:

• The KD-trained student model achieved high accuracy, with a Gamma passing rate (2%/2 mm) of 98.13 ± 0.76% against ground truth.
• Inference times were significantly reduced: 11 ms on GPU and 374 ms on CPU for the student model.
• The compact engine demonstrated effective conversion of low-accuracy doses to high-accuracy patient doses.

Conclusions:

• The KD framework enables the creation of compact dose engines with accuracy comparable to larger models.
• This approach significantly lowers computational burdens and hardware requirements, enhancing clinical applicability.
• The developed engine offers a promising solution for efficient and accurate dose calculation in VMAT radiotherapy.