Dose distribution prediction in isodose feature-preserving voxelization domain using deep convolutional neural...

Area of Science:

• Medical Physics
• Radiotherapy
• Computational Biology

Background:

• Accurate dose prediction is crucial for effective radiotherapy planning.
• Conventional voxel-based methods can be computationally intensive.
• Simplifying dose distribution representation is key to improving efficiency.

Purpose of the Study:

• To implement a deep convolutional neural network (CNN) framework for dose prediction.
• To utilize the concept of isodose feature-preserving voxelization (IFPV) for simplified dose distribution representation.
• To evaluate the performance of the IFPV-based CNN model in predicting dose distributions.

Main Methods:

• Developed a deep CNN model trained on 60 volumetric modulated arc therapy (VMAT) treatment plans.
• Employed isodose feature-preserving voxelization (IFPV) for concise representation of dose distributions.
• Validated the model on 10 independent prostate cancer cases using DVH comparison, dose difference maps, and sum of absolute residual (SAR).

Main Results:

• The IFPV-based CNN method demonstrated good prediction performance comparable to conventional voxel-based methods.
• Achieved mean SARs of 0.029 ± 0.020 for bladder and 0.077 ± 0.030 for rectum.
• Significantly reduced the number of dose representation points while maintaining accuracy.

Conclusions:

• A novel deep CNN-based dose prediction method using IFPV was successfully developed.
• The IFPV approach offers a promising strategy for enhancing dose prediction efficiency.
• This method has the potential to streamline the radiotherapy treatment planning workflow.