A convolution neural network for higher resolution dose prediction in prostate volumetric modulated arc therapy

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence

Background:

• Accurate dose calculation is crucial for effective radiotherapy.
• Low-resolution dose distributions are computationally efficient but lack precision.
• Deep learning offers potential for enhancing dose prediction accuracy.

Purpose of the Study:

• To assess the feasibility of using convolutional neural networks (CNNs) to predict high-resolution dose distributions from low-resolution inputs.
• To compare the performance of a U-net model using only low-resolution dose data versus one using both dose and CT data.

Main Methods:

• A U-net model was employed to predict 2 mm grid dose distributions from 5 mm grid inputs for prostate cancer VMAT plans.
• Two models were investigated: one using only low-resolution dose (D model) and another incorporating CT data (DC model).
• Dice Similarity Coefficient (DSC) and gamma analysis were used for evaluation against reference dose distributions.

Main Results:

• The DC model demonstrated significantly higher DSC values compared to the D model (p < 0.01).
• Gamma passing rates for CTV, PTV, and bladder were significantly higher in the DC model (p < 0.002-0.02).
• Mean doses for CTV and PTV in the DC model showed significantly better agreement with reference doses (p < 0.0001).

Conclusions:

• The U-net model incorporating both dose and CT image data as input significantly enhances the accuracy of predicted dose distributions.
• This approach shows promise for improving dose prediction in radiotherapy planning.