Decompose kV projection using neural network for improved motion tracking in paraspinal SBRT.

Area of Science:

• Medical Physics
• Radiotherapy Technology
• Artificial Intelligence in Medicine

Background:

• Markerless motion tracking in stereotactic body radiation therapy (SBRT) for paraspinal tumors faces challenges with low contrast in on-treatment kV images, especially from lateral angles.
• Overlapping spine and surrounding organs in X-ray images can lead to auto-registration errors, compromising accurate motion management.

Purpose of the Study:

• To develop an automated method for extracting spine components from 2D kV X-ray images.
• To enhance the accuracy and robustness of motion management in paraspinal SBRT through improved spine visualization.

Main Methods:

• A ResNet generative adversarial network (ResNetGAN) was employed to learn the transformation from 2D kV images to reference spine digitally reconstructed radiographs (DRRs).
• A specialized multi-channel, multi-domain loss function was utilized to optimize the quality of the generated spine images.
• The model was trained on 1347 kV images and evaluated on 226 kV images, comparing registration accuracy against reference spine DRRs.

Main Results:

• The generated spine images demonstrated high fidelity, with mean Peak Signal-to-Noise Ratio (PSNR) of 60.08 and Structural Similarity Index Measure (SSIM) of 0.99.
• Submillimeter accuracy in spine tracking was achieved, with mean errors of 0.13 mm and 0.12 mm in the $$ and $$ directions, respectively.
• The accuracy improvements were consistent across various X-ray beam angles, including lateral and anteroposterior projections.

Conclusions:

• A deep learning-based approach effectively removes soft tissue information from kV images.
• This method significantly enhances the accuracy of spine tracking for paraspinal SBRT.
• The developed technique offers a more robust solution for motion management in image-guided radiation therapy.