Application of Recursive Convolutional Layers to a 3D U-Net for Microbeam Radiotherapy dose prediction

Area of Science:

• Medical Physics
• Radiotherapy
• Computational Biology

Background:

• Microbeam Radiotherapy (MRT) is a spatially fractionated radiotherapy technique creating high and low dose areas.
• 3D U-Nets can accelerate Monte Carlo dose calculations for MRT but may lack accuracy, especially in low-dose (valley) regions.
• Recursive Convolutional Layers (RCLs) are known to improve neural network performance but haven't been applied to MRT dose prediction.

Purpose of the Study:

• To apply RCLs to improve the accuracy of predicting Geant4-simulated MRT doses in rats.
• To investigate the impact of RCLs on both high-dose (peak) and low-dose (valley) regions in MRT dose prediction.

Main Methods:

• An existing 3D U-Net architecture for MRT was modified to incorporate RCLs with five recursions.
• The effect of RCLs was evaluated at various network locations for both peak and valley dose predictions.
• Parts of the 3D U-Net were replaced with sequential RCL blocks to assess the impact on virtual network depth.

Main Results:

• RCLs significantly improved the accuracy of both peak and valley dose predictions.
• Valley dose prediction accuracy saw the most substantial improvement, increasing from 70.6% to 84.9% agreement with Geant4 simulations.
• Optimal placement for RCLs was found to be immediately after the input and before the output layers of the network.

Conclusions:

• Incorporating RCLs into 3D U-Nets is an effective strategy for enhancing MRT dose prediction accuracy.
• The improvements are most notable in predicting low-dose regions, addressing a key limitation of previous U-Net models.
• This approach offers a promising method for more accurate and efficient MRT dose calculations.