ESTRO 2024 - Abstract Book

S3964

Physics - Image acquisition and processing

ESTRO 2024

Results:

Figures 1 and 2 compare cone-beam CT images from the Gamma Knife ICON system, generated synthetic CBCT images, and planning CT scans for a test patient. Using the region of interest method, the model achieved an optimal SSIM score of 0.93 on test datasets, signifying a high resemblance between the synthetic CBCT and the ground truth planning CT images. However, field-of-view limitations in CBCT images (as depicted in Figure 1a) tend to lower the SSIM when compared to ground truth CT images for the actual field-of-view. The use of SELU activation in the generator outperformed the original leaky ReLU in the U-Net encoder and decoder. Figure 3 shows the comparison of the Hounsfield units between CBCT, synthetic CBCT, and planning CT.