Spatial feature and resolution maximization GAN for bone suppression in chest radiographs.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Chest radiographs (CXRs) are vital for lung pathology visualization.
• Bone structures in CXRs obscure lesions, hindering accurate diagnosis.
• Existing bone suppression methods often compromise image quality and spatial details.

Purpose of the Study:

• To develop an advanced Generative Adversarial Network (GAN) for effective bone suppression in CXRs.
• To enhance the retention of spatial features and overall image quality during bone removal.
• To improve the diagnostic utility of CXRs by minimizing bone interference.

Main Methods:

• Modified pix2pix GAN architecture incorporating Wasserstein GAN with Gradient Penalty for improved stability.
• Utilized a combination of L1, Perceptual, and Sobel loss functions in the generator for comprehensive information capture.
• Employed specific hyperparameters (δ=10^4, α=1, β=10, γ=10) for optimal bone suppression and quality retention.

Main Results:

• Achieved a mean PSNR of 43.588, NMSE of 0.00025, SSIM of 0.989, and Entropy of 0.454 bits/pixel.
• Demonstrated significant improvement in bone suppression and spatial information preservation.
• The proposed SFRM-GAN outperformed state-of-the-art models in denoising and image quality retention.

Conclusions:

• The integration of Sobel and Perceptual loss functions enhances bone suppression and spatial information preservation.
• SFRM-GAN effectively suppresses bones while maintaining critical image quality and intrinsic information.
• The model provides statistically significant results and is suitable for medical image denoising and preprocessing.