Feasibility of photon beam profile deconvolution using a neural network.

Area of Science:

• Medical Physics
• Radiotherapy Physics
• Computational Physics

Background:

• Ionization chambers are standard for photon beam profile scanning.
• Volume averaging effect (VAE) in ionization chambers artificially broadens penumbra width by 2-3 mm.
• Accurate beam profiles are crucial for precise radiotherapy dose calculations.

Purpose of the Study:

• To investigate the feasibility of photon beam profile deconvolution using a three-layer feedforward neural network.
• To eliminate the Volume Averaging Effect (VAE) from ionization chamber-measured beam profiles.
• To improve the accuracy of photon beam profile measurements in radiotherapy.

Main Methods:

• Collected transverse photon beam profiles using ionization chambers and diode detectors.
• Trained a three-layer feedforward neural network with ionization chamber data.
• Utilized a sliding window approach for input data extraction and deconvolution.
• Optimized network parameters (window width, hidden neurons) via parametric sweeping.
• Evaluated agreement using Mean Squared Error (MSE) and Penumbra Width Difference (PWD).

Main Results:

• Selected optimal network parameters: 15 input neurons (window width) and 5 hidden neurons.
• Achieved excellent agreement between deconvolved and reference (diode) profiles across all datasets.
• Significantly reduced mean Penumbra Width Difference (PWD) from ~2.4 mm to ~0.1 mm post-deconvolution.
• Demonstrated successful generalization ability of the trained neural network on unseen test data.

Conclusions:

• Validated the feasibility of using feedforward neural networks for photon beam profile deconvolution.
• The three-layer neural network effectively eliminated VAE, yielding profiles in excellent agreement with diode measurements.
• This AI-driven approach offers a promising solution for enhancing the accuracy of radiotherapy beam characterization.