Fast dose calculation in x-ray guided interventions by using deep learning.

Area of Science:

• Medical Physics
• Radiology
• Artificial Intelligence in Medicine

Background:

• Accurate patient dose estimation is critical for preventing radiation-induced side effects during X-ray guided interventions.
• Current methods using reference air kerma lack precision due to patient-specific morphology and organ composition.
• Monte Carlo (MC) simulations offer accuracy but are computationally intensive, limiting intra-operative use.

Purpose of the Study:

• To develop a rapid deep convolutional neural network (CNN) for precise patient dose estimation during X-ray guided interventions.
• To create personalized 3D dose maps using MC simulations and a modified 3D U-Net architecture.
• To validate the CNN's accuracy against clinical measurements and assess its potential for real-time dose monitoring.

Main Methods:

• A modified 3D U-Net CNN was trained using MC simulations of X-ray irradiation on 82 patient CT scans.
• Simulations incorporated variations in X-ray source angulation, position, and tube voltage for abdominal procedures.
• Clinical validation involved comparing MC-derived dose maps with skin dose measurements during endovascular abdominal aortic repairs.

Main Results:

• Clinical validation showed an average error of 5.1% at specific anatomical points.
• The CNN achieved test errors of 11.5 ± 4.6% for peak skin dose and 6.2 ± 1.5% for average skin dose.
• Mean errors for abdominal region and pancreas doses were 5.0 ± 1.4% and 13.1 ± 2.7%, respectively.

Conclusions:

• The developed deep learning network accurately predicts personalized 3D dose maps based on patient CT scans and imaging parameters.
• The network's rapid computation time makes it suitable for intra-operative dose monitoring and reporting.
• This AI-driven approach offers a significant advancement for radiation safety in interventional radiology.