Deep learning based de-overlapping correction of projections from a flat-panel micro array X-ray source: Simulation...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Flat-panel X-ray sources offer space and time savings for static computer tomography (CT).
• Overlapping cone beams from micro-ray sources cause significant structural blurring in projection results.
• Traditional deoverlapping methods are insufficient for addressing these issues.

Purpose of the Study:

• To develop and validate a novel method for deblurring and deoverlapping X-ray cone beam projections.
• To improve image quality in static CT applications using flat-panel X-ray sources.

Main Methods:

• A U-like neural network was employed to convert overlapping cone beam projections into parallel beam projections.
• Structural Similarity (SSIM) loss was utilized as the primary loss function during training.
• The model was trained and tested on Shepp-Logan, line-pairs, and abdominal datasets, with performance evaluated using MSE, PSNR, and SSIM.

Main Results:

• The U-net model achieved excellent results, including a SSIM of 0.998 for the Shepp-Logan dataset.
• For abdominal data, the model yielded MSE of 1.563×10⁻³, PSNR of 28.0586 dB, and SSIM of 0.983.
• The model demonstrated good generalization capabilities on unseen head phantom data.

Conclusions:

• The end-to-end U-net architecture is feasible for deblurring and deoverlapping in flat-panel X-ray imaging.
• This approach offers a promising solution for enhancing image clarity in CT applications.
• The study validates the effectiveness of the proposed neural network for overcoming projection overlap challenges.