Machine Learning-Based X-Ray Projection Interpolation for Improved 4D-CBCT Reconstruction

Area of Science:

• Medical Imaging
• Radiology
• Computer Vision

Background:

• Respiration-correlated cone-beam computed tomography (4D-CBCT) generates dynamic volumetric images but is limited by projection data quality.
• Image quality is directly influenced by the number of available CBCT projections for reconstruction.
• Interpolation techniques can create intermediate projections to improve reconstruction.

Purpose of the Study:

• To investigate the use of transfer learning and novel regression models for generating intermediate projections in 4D-CBCT.
• To evaluate the performance of state-of-the-art deep learning video frame interpolation models for this task.
• To assess the impact of interpolated projections on the final 4D-CBCT image quality.

Main Methods:

• Employed pre-trained deep learning models, including the Real-Time Intermediate Flow Estimation (RIFE) algorithm, for video frame interpolation.
• Developed a novel regression predictive modeling approach to generate intermediary projections.
• Validated model performance using digital phantom and clinical datasets.

Main Results:

• The RIFE algorithm demonstrated superior performance, achieving high SSIM (0.986 ± 0.010), PSNR (44.13 ± 2.76), and low MSE (18.86 ± 206.90).
• 4D-CBCT images reconstructed with interpolated projections showed reduced streaking artifacts compared to those reconstructed with original projections alone.
• Transfer learning algorithms effectively enhanced 4D-CBCT image quality.

Conclusions:

• Transfer learning, particularly using models like RIFE, offers a significant advantage for improving 4D-CBCT image quality.
• The proposed methods successfully generated intermediate projections, leading to enhanced image clarity and artifact reduction.
• This approach holds promise for improving diagnostic accuracy in 4D-CBCT imaging.