Personalized brachytherapy dose reconstruction using deep learning.

Area of Science:

• Medical Physics
• Radiation Oncology
• Computational Biology

Background:

• Brachytherapy offers a superior treatment gain factor over external radiation therapy due to accurate dose delivery.
• Traditional dosimetry models like TG-43 simplify calculations by ignoring medium heterogeneities, potentially impacting dose accuracy.
• Accurate absorbed dose calculation is crucial for optimizing tumor control and minimizing normal tissue toxicity in brachytherapy.

Purpose of the Study:

• To develop a deep learning (DL) model for accurate, patient-specific brachytherapy dosimetry.
• To overcome the limitations of TG-43 by incorporating dosimetric impacts of medium heterogeneities.
• To achieve accurate dose distributions within a reasonable computation time.

Main Methods:

• A Monte Carlo (MC)-based personalized brachytherapy dosimetry simulator (PBrDoseSim) was developed.
• A deep neural network (DNN) was trained using MC simulations as ground truth.
• Input data included dose distribution kernels in water and CT-derived volumetric density maps to account for heterogeneity.

Main Results:

• The DNN-predicted dose kernels showed excellent agreement with MC-based kernels (MRAE: 1.16 ± 0.42%).
• Dose volume histogram (DVH) parameters (D90, V150, V100) in the clinical target volume demonstrated high accuracy (MRAE: 1.8 ± 0.86% for D90).
• Organ-at-risk (bladder, sigmoid, rectum) dose calculations also showed good agreement with MC (MRAE: 2.7 ± 1.7% for bladder D5cc).

Conclusions:

• The proposed DNN-based approach achieves dosimetry comparable to MC methods.
• This DL approach effectively addresses the computational burden of MC simulations.
• The method overcomes the oversimplifications inherent in TG-43, enabling more accurate personalized brachytherapy.