Prediction of dose deposition matrix using voxel features driven machine learning approach

Area of Science:

• Medical Physics
• Computational Biology
• Radiotherapy

Background:

• Radiation therapy planning requires accurate dose calculation.
• Current methods like Monte Carlo (MC) are computationally intensive.
• Pencil Beam (PB) algorithms offer speed but lack precision.

Purpose of the Study:

• To develop a machine learning (ML) based method for predicting dose deposition matrices (DDM).
• To improve the speed and accuracy of radiation therapy plan optimization.

Main Methods:

• A cascade forward backprop neural network was trained using voxel features.
• Features included distances to beamlet axis, density, and PB dose.
• The ML-predicted DDM was used for plan optimization and compared to MC and PB methods.

Main Results:

• The ML method achieved significantly lower Mean Absolute Error (MAE) compared to the PB method for both head and lung tumors.
• For head tumors, ML MAE was 0.49 × 10^-4 vs. PB MAE of 1.86 × 10^-4.
• For lung tumors, ML MAE was 1.42 × 10^-4 vs. PB MAE of 3.72 × 10^-4.
• ML method's PTV dose coverage (D₉₈) was within 1.2% (head) and 2.1% (lung) of MC, while PB differed by over 10% and 16% respectively.

Conclusions:

• A reliable ML-based DDM prediction method was established for radiation therapy plan optimization.
• The ML method provides plans comparable to MC and superior to PB in accuracy.
• This approach facilitates rapid plan optimization and accurate dose calculation.