Deep learning-based projection synthesis for low-dose cone-beam computed tomography imaging in image-guided...

Area of Science:

• Medical Imaging
• Radiotherapy Technology
• Artificial Intelligence in Medicine

Background:

• Cone-beam computed tomography (CBCT) in image-guided radiotherapy (IGRT) involves imaging doses that can adversely affect patient health.
• Improving the quality of sparse-view, low-dose CBCT images is crucial for safe and effective radiotherapy.

Purpose of the Study:

• To introduce and evaluate a projection synthesis convolutional neural network (SynCNN) model for enhancing sparse-view low-dose CBCT images.
• To assess the SynCNN model's effectiveness in synthesizing missing projections and reconstructing high-quality tomographic images.

Main Methods:

• A retrospective study involving 223 brain tumor patients was conducted.
• The SynCNN model utilized direction-separable spatial kernels to synthesize missing projections from neighboring sparse-view data.
• Image quality was evaluated using metrics like RMSE, PSNR, SSIM, and expert subjective grading, with comparisons against other methods.

Main Results:

• The SynCNN model accurately synthesized missing projections, significantly improving image quality in both phantom and patient studies.
• Quantitative metrics (RMSE, PSNR, SSIM) showed substantial improvements with SynCNN synthesis across various sparse-view rates (1/2, 1/4, 1/8).
• Expert evaluations confirmed higher image quality and acceptance rates for SynCNN-synthesized images, with synthesis time under 0.01s.

Conclusions:

• The SynCNN model effectively enhances sparse-view CBCT image quality with minimal computational cost.
• This technology holds potential for reducing CBCT imaging dose in IGRT, thereby improving patient safety.