A Variational Network for Biomedical Images Denoising using Bayesian model and Auto-Encoder.

Area of Science:

• Biomedical imaging
• Computer vision
• Deep learning

Background:

• Traditional autoencoders and CNNs for biomedical image denoising require training on known noise, limiting generalization to new noise distributions.
• Current methods often fail to accurately denoise images with unknown or varying noise characteristics.

Purpose of the Study:

• To propose a novel variational network for biomedical image denoising using a Bayesian approach.
• To develop a method that effectively denoises images with consistent noise distributions.
• To improve the accuracy and reliability of biomedical image analysis for clinical applications.

Main Methods:

• A Bayesian approach was used to estimate noise distributions by calculating posterior distributions.
• A variational network was trained using a loss function combining Bayesian prior and autoencoder objectives.
• The method was tested on CT-Scan datasets and compared against state-of-the-art denoising techniques.

Main Results:

• The proposed method demonstrated superior denoising accuracy and visual quality compared to existing techniques.
• Achieved a Peak Signal-to-Noise Ratio (PSNR) of 39.18 dB and a Structural Similarity Index Measure (SSIM) of 0.9941 for noise intensity std = 10.
• Showcased improved computational efficiency in denoising biomedical images.

Conclusions:

• The integration of Bayesian modeling and variational networks offers an effective solution for biomedical image denoising.
• This approach has the potential to significantly enhance clinical diagnosis and treatment planning through improved image analysis.