A Comparison Study Between CNN-Based Deformed Planning CT and CycleGAN-Based Synthetic CT Methods for Improving iCBCT...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Cone-Beam CT (CBCT)

Background:

• Improving image quality in cone-beam CT (CBCT) is crucial for accurate radiation therapy planning.
• Varian Halcyon system's in-room CBCT (iCBCT) requires enhancement for optimal diagnostic and treatment verification purposes.

Purpose of the Study:

• To compare the effectiveness of two image quality improvement methods for Varian Halcyon iCBCT.
• To evaluate deformed planning CT (dpCT) via convolutional neural network (CNN) against synthetic CT (sCT) generation via CycleGAN.

Main Methods:

• Utilized 190 paired pelvic CT and iCBCT datasets for training and testing.
• Developed a 3D multi-stage registration network (MSnet) for dpCT and trained a 2D CycleGAN for sCT generation.
• Assessed registration and sCT accuracy using Dice Similarity Coefficient (DSC) on organs like bladder, spinal cord, and bone marrow.

Main Results:

• Both dpCT and sCT methods improved iCBCT image quality.
• sCT generation achieved higher accuracy (DSC 0.884 vs. 0.769 for bladder) compared to dpCT.
• sCT demonstrated significantly better performance when planning CT and iCBCT volume differences were substantial (DSC 0.859 vs. 0.596 for bladder).

Conclusions:

• Both registration and sCT generation are effective in enhancing iCBCT image quality.
• sCT generation offers superior accuracy, particularly in cases with significant anatomical variations between planning CT and iCBCT.