Minimum imaging dose for deep learning-based pelvic synthetic computed tomography generation from cone beam images

Area of Science:

• Medical Physics
• Radiotherapy Imaging
• Artificial Intelligence in Medicine

Background:

• Daily cone-beam computed tomography (CBCT) in image-guided radiotherapy (IGRT) involves radiation exposure and secondary cancer risk.
• Reducing imaging dose is challenging due to image quality deterioration.

Purpose of the Study:

• Investigate reduced imaging dose levels for CBCT by decreasing projections.
• Evaluate deep learning algorithms for image correction to identify the lowest achievable imaging dose while maintaining quality.

Main Methods:

• CBCT images were reconstructed using 100%, 25%, 15%, and 10% of projections.
• Two deep learning models, cycleGAN and CUT, were trained and compared for generating synthetic CT (sCT) from low-dose CBCT.
• Image quality, Hounsfield unit (HU) accuracy, positioning accuracy, and anatomical segmentation (bladder, rectum) were evaluated.

Main Results:

• All synthetic CTs (sCTs) achieved high image quality metrics (MAE < 59 HU, SSIM > 0.94, PSNR > 33 dB).
• Dosimetric and positioning accuracies were maintained within acceptable limits (DVH differences < 2 Gy, positioning < 0.30 mm/0.30°).
• CycleGAN demonstrated superior performance in anatomical segmentation (Dice scores 0.85/0.81 for bladder/rectum) compared to CUT (0.83/0.76), though both showed decreased accuracy at 15% and 10% projections.

Conclusions:

• Deep learning algorithms effectively generate synthetic CTs from reduced CBCT imaging doses.
• Based on segmentation accuracy, 25% projection is identified as the minimum acceptable imaging dose for CBCT in IGRT.