Geometry-encoded deep learning (GeoDL) framework for real-time 3D dose verification for online adaptive radiotherapy

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Online adaptive radiotherapy (ART) requires fast and accurate dose verification.
• Current methods are computationally intensive, hindering ART workflow efficiency.
• Previous work showed potential for 2D dose estimation using geometry-encoded deep learning (GeoDL).

Purpose of the Study:

• To investigate the feasibility of extending GeoDL for real-time 3D dose verification.
• To address the unmet clinical need for rapid dose assessment during ART.
• To evaluate GeoDL's performance across the entire 3D volume of interest.

Main Methods:

• Developed a GeoDL algorithm unifying fluence maps and CT images using linear accelerator geometry.
• Employed a 3D U-Net architecture trained on integrated CT-FM volumes for 3D dose estimation.
• Utilized a combined loss function to optimize predictions in high and low dose regions; trained and tested on prostate cancer cases.

Main Results:

• Achieved highly accurate 3D dose estimation in near real-time (~35ms per case).
• Demonstrated excellent dosimetric accuracy with an average γ passing rate (3%/2mm) of 99.94% ± 0.15%.
• Reported minimal mean dose differences for planning target volumes (PTV) and organs at risk (OARs).

Conclusions:

• GeoDL is feasible for ultra-fast dose verification in ART.
• The algorithm shows strong performance in key dosimetric metrics for the entire 3D volume.
• GeoDL is a promising tool for real-world clinical application in radiotherapy.