Deep learning-based 3D dose reconstruction for intensity modulated radiation therapy using electronic portal imaging...

Area of Science:

• Medical Physics
• Radiotherapy Physics
• Medical Imaging

Background:

• Electronic portal imaging devices (EPIDs) are used for in vivo dosimetry to detect treatment discrepancies.
• Accurate 3D dose reconstruction is crucial for radiotherapy quality assurance.

Purpose of the Study:

• To develop and evaluate a deep learning framework for direct 3D patient dose reconstruction from EPID images and planning CT.
• To avoid reliance on Monte Carlo simulations or conventional back-projection methods.

Main Methods:

• A Res-UNet architecture was trained on head and neck IMRT patient data.
• Dose slices were predicted from EPID and CT data, assembled into 3D volumes, and summed across beams.
• Evaluation involved voxel-wise error metrics, 3D gamma analysis, and DVH comparisons.

Main Results:

• The framework achieved a mean voxel-wise MAE of 0.163 Gy, with >95% of voxels < 2 Gy.
• Overall 3D gamma passing rates were 90.67% for the 3%/3 mm criterion.
• Most DVH comparisons showed no significant differences, though some organs-at-risk did.

Conclusions:

• The deep learning framework enables fast, patient-specific 3D dose reconstruction without complex physics modeling.
• The method demonstrates feasibility for quality assurance and potential for online adaptive radiotherapy.
• High-gradient regions require further attention, but accuracy is sufficient for many applications.