Feasibility of two-dimensional dose distribution deconvolution using convolution neural networks.

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning in Healthcare

Background:

• Accurate dose distribution is critical for effective radiotherapy.
• Scintillation detectors can introduce artifacts, particularly in the penumbra region.
• Traditional analytical methods for dose deconvolution have limitations.

Purpose of the Study:

• To evaluate the feasibility of using convolutional neural networks (CNNs) for two-dimensional (2D) dose distribution deconvolution.
• To address detector-interface artifacts in the penumbra region of an in-house scintillation detector.
• To compare CNN-based deconvolution with analytical approaches.

Main Methods:

• Developed a shallow linear CNN, PenumbraNet, for dose deconvolution.
• Trained PenumbraNet on datasets from a Novalis Tx medical linear accelerator, including square fields and IMRT plans.
• Validated PenumbraNet using test data and evaluated performance with gamma index criteria (2%/2mm, 3%/3mm), comparing with analytical Wiener filtering and different CNN architectures.

Main Results:

• PenumbraNet achieved a mean gamma passing rate of 95.81% (3%/3mm), significantly outperforming analytical Wiener filtering (84.77%).
• Nonlinear PenumbraNet variants showed even higher performance, with the best achieving 96.62% (3%/3mm).
• The best performing nonlinear PenumbraNet improved the gamma passing rate by 11.85% compared to the measured dose distribution.

Conclusions:

• Demonstrated the feasibility and effectiveness of PenumbraNets for 2D dose distribution deconvolution.
• CNN-based deconvolution, particularly nonlinear variants, offers superior accuracy in correcting for detector artifacts.
• This approach holds potential for enhancing dose calculation accuracy in clinical radiotherapy.