Is deep learning-enabled real-time personalized CT dosimetry feasible using only patient images as input?

Area of Science:

• Medical Physics
• Radiology
• Artificial Intelligence

Background:

• Personalized CT dosimetry is crucial for radiation safety but is currently labor-intensive.
• Existing methods rely on time-consuming Monte Carlo simulations for accurate organ dose estimation.
• There is a need for faster, more efficient dosimetry techniques in clinical practice.

Purpose of the Study:

• To introduce a novel deep learning-based dosimetry method for rapid and precise organ dose estimation.
• To utilize only patient computed tomography (CT) images as input for the proposed dosimetry approach.
• To overcome the limitations of traditional, time-intensive dosimetry methods.

Main Methods:

• Employed conditional generative adversarial networks (cGANs) for image-to-image translation to generate synthetic dose images.
• Utilized the pix2pix architecture combined with a regression model for dose image synthesis.
• Performed manual segmentation of organs (lungs, heart, breast, bone, skin) to compare dose calculations.

Main Results:

• Achieved an average organ dose estimation error of 8.3%, with a maximum error of 20% across all considered organs.
• Demonstrated the method's clinical feasibility with an automated organ dose calculation pipeline.
• Calculated organ doses for the heart and lungs per CT slice in approximately 2 seconds.

Conclusions:

• Deep learning enables real-time personalized CT dosimetry.
• The proposed method accurately estimates organ doses using only patient CT images.
• This approach offers a feasible solution for rapid, individualized dosimetry in clinical settings.