Rapid estimation of patient-specific organ doses using a deep learning network.

Area of Science:

• Medical Imaging
• Radiological Physics
• Artificial Intelligence in Medicine

Background:

• Accurate patient-specific organ dose estimation in diagnostic CT is crucial for assessing secondary cancer risks and optimizing protocols.
• Current methods, such as Monte Carlo simulations, are time-consuming and often rely on generalized phantoms.

Purpose of the Study:

• To develop and validate a rapid workflow utilizing deep learning networks for estimating individual organ doses from thorax CT scans.
• To demonstrate the feasibility of near real-time organ dose estimation for improved patient management.

Main Methods:

• A deep learning network was trained using Monte Carlo-derived 3D dose distributions from 95 patients undergoing thorax CT.
• Independent variables included water-equivalent diameter, scan length, and tube current.
• The model predicted doses for the heart, lungs, esophagus, and bones, and was validated on a separate dataset of 19 patients.

Main Results:

• The deep learning networks provided organ dose predictions within one second.
• Predictions showed excellent agreement with Monte Carlo simulations, with average differences ranging from -5.1% to 4.3% across organs and datasets.
• The workflow achieved high accuracy for patient-specific organ dose estimation.

Conclusions:

• Patient-specific organ doses can be estimated in nearly real-time using the proposed deep learning workflow.
• This approach is readily implementable and requires minimal training data.
• The method holds significant potential for enhancing radiation safety and personalized medicine in CT imaging.