PIDA-GAN: physics-informed dual-stage attention GAN with data consistency for sparse-view CT reconstruction

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Sparse-view CT (SVCT) reconstruction is challenging due to insufficient data, causing noise and artifacts.
• Existing deep learning methods often oversmooth details and lack data consistency.
• A physics-informed approach is needed to improve SVCT reconstruction accuracy and realism.

Purpose of the Study:

• To develop a physics-informed dual-stage generative adversarial framework for SVCT reconstruction.
• To enforce data consistency and preserve fine anatomical structures.
• To improve perceptual realism in SVCT images.

Main Methods:

• Proposed PIDA-GAN, a dual-stage attention GAN with U-Net generators, self-attention, and adaptive gating.
• Incorporated a physics-informed data consistency layer with iterative updates using forward/backprojection.
• Employed an edge-aware discriminator and a hybrid loss function (pixel-wise, SSIM, perceptual, adversarial).

Main Results:

• PIDA-GAN outperformed state-of-the-art methods in PSNR, SSIM, and RMSE across various view settings (60, 30, 16, 4).
• Demonstrated significant improvements under extreme sparsity, effectively removing artifacts and reducing oversmoothing.
• Showcased strong cross-dataset generalization on the LIDC-IDRI dataset without fine-tuning.

Conclusions:

• PIDA-GAN offers a robust, physics-informed framework for ultra-sparse-view CT reconstruction.
• The method effectively preserves structural fidelity and perceptual realism.
• PIDA-GAN holds significant potential for clinical low-dose CT applications.