Deep learning-based upsampling of 2D detector array measurements for patient plan verification in radiotherapy

Area of Science:

• Medical Physics
• Radiotherapy
• Image Processing

Background:

• Detector arrays are crucial for intensity modulated radiation therapy (IMRT) plan verification.
• Limited intrinsic resolution of detector arrays can lead to inaccuracies in steep dose gradients and narrow peaks.

Purpose of the Study:

• To develop a deep learning approach for enhancing the spatial resolution of detector arrays used in patient plan verification.
• To augment missing data and increase sampling frequency of 2D dose profiles, correcting for volume-averaging effects.

Main Methods:

• Utilized Monte Carlo simulations to generate synthetic training data for various linear accelerator setups and field shapes.
• Implemented a deep convolutional neural network (CNN) architecture trained with PyTorch for the OCTAVIUS Detector 1500.
• Validated the approach by comparing upsampled measurements with high-resolution detector arrays and radiochromic film, and reconstructed 3D dose distributions with treatment planning system (TPS) calculations.

Main Results:

• Achieved a threefold increase in resolution from 5 mm to 1.7 mm using neural networks.
• Demonstrated an average increase in gamma index passing rate of up to 20% for IMRT segments compared to bilinear interpolation.
• Showed average passing rate increases of 22% for 3D dose reconstructions and 7-8% for VMAT plans.

Conclusions:

• Neural network upsampling effectively enhances detector array resolution for radiation therapy.
• The method provides superior interpolation of measurement points, particularly in steep gradient regions, compared to standard bilinear interpolation.
• Application of the neural network upsampling approach consistently increases passing rates for investigated VMAT plans.