Metric learning guided sinogram denoising for cone beam CT enhancement

Area of Science:

• Medical Imaging
• Computational Imaging
• Image Reconstruction

Background:

• Cone beam computed tomography (CBCT) is widely used but limited by low detail conspicuity and quantitative accuracy.
• Traditional post-reconstruction denoising methods often introduce artifacts and are computationally intensive.
• Joint denoising-reconstruction is complex and computationally expensive, hindering clinical utility.

Purpose of the Study:

• To develop and evaluate a novel Metric-learning guided wavelet transform reconstruction (MEGATRON) approach.
• To enhance CBCT image quality using projection-domain processing for improved clinical utility.
• To overcome the limitations of traditional denoising and reconstruction techniques.

Main Methods:

• Developed a metric learning module to define enhancement objectives in the wavelet encoded sinogram domain.
• Employed a denoising network (res-Unet) to map measured cone-beam projections to enhanced versions.
• Translated image-domain quality enhancement goals to the projection domain for efficient processing.
• Utilized simulated projections from the X-ray based Cancer Imaging Simulation Toolkit (XCIST) and a dataset of 123 human thoraxes from OSIC.

Main Results:

• MEGATRON achieved quantitative performance metrics including RMSE of 30.97 ± 4.25 HU, PSNR of 37.45 ± 1.78, and SSIM of 93.23 ± 1.62.
• These results are comparable to sophisticated physics-driven CBCT enhancement methods.
• Demonstrated the promise and utility of the MEGATRON approach in improving CBCT image quality.

Conclusions:

• Incorporating metric learning into sinogram denoising enhances reconstruction goals and quantitative performance.
• The proposed MEGATRON approach is compatible with various denoiser network structures and reconstruction modules.
• This method offers a flexible and effective solution for improving CBCT image quality and clinical applicability.