Synthetic CT generation from cone-beam CT using deep learning for head-and-neck adaptive radiotherapy

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Synthetic computed tomography (CT) generation from cone-beam CT (CBCT) is crucial for adaptive radiotherapy (ART) on conventional linear accelerators.
• Clinical implementation of synthetic CT generation is currently limited.

Purpose of the Study:

• To develop and validate a deep learning model for synthetic CT generation in head-and-neck cancer patients.
• To evaluate the model's accuracy for adaptive radiotherapy workflows, including regions outside the CBCT field of view.

Main Methods:

• A UNet3+ generative adversarial network with multi-component losses was trained on 148 head-and-neck cancer patients.
• Model performance was assessed using anatomical concordance, intensity fidelity against planning CT, and dose recalculation accuracy (gamma index).

Main Results:

• The model achieved a global mean absolute error of 49.9 HU, with lower error within the CBCT field of view.
• Dose recalculation showed mean differences <1.6% in planning target volume dose metrics and high gamma pass rates (99.0% at 3%/2mm).
• Surface Dice similarity coefficient was 0.81 at a 2mm tolerance, indicating good anatomical agreement.

Conclusions:

• The developed deep learning model accurately generates synthetic CT from CBCT, including regions beyond the CBCT field of view.
• The model supports the feasibility of daily adaptive radiotherapy dose recalculation on conventional linear accelerators.
• Prospective clinical evaluation is warranted to further validate the model's utility.