Deep image and feature prior algorithm based on U-ConformerNet structure.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Deep Image Prior (DIP) is a powerful technique for image reconstruction.
• Conventional DIP methods using convolutional layers have limitations in performance and artifact suppression.
• Enhancing DIP's potential requires novel network architectures and regularization strategies.

Purpose of the Study:

• To improve the image reconstruction quality and artifact suppression of the Deep Image Prior (DIP) approach.
• To introduce a superior DIP algorithm that overcomes the limitations of traditional convolutional structures.
• To validate the enhanced performance in a real-world low-dose CT scenario.

Main Methods:

• Developed a novel U-ConformerNet architecture to replace traditional U-Net convolutional layers in DIP.
• Incorporated the 'lpips' deep network-based feature distance regularization method.
• Enabled flexible switching between supervised and unsupervised learning modes.

Main Results:

• Achieved a Peak Signal-to-Noise Ratio (PSNR) >35 dB in unsupervised mode and >36 dB in supervised mode on the LoDoPaB dataset.
• Demonstrated superior performance compared to the DIP-TV algorithm in both noise eradication and artifact suppression.
• Confirmed a positive correlation between method accuracy and the quality of the a priori image, leveraging deep networks.

Conclusions:

• The U-ConformerNet-based DIP algorithm improves unsupervised PSNR by 2-3 dB over DIP-TV.
• In supervised mode, the proposed algorithm achieves performance comparable to state-of-the-art end-to-end deep learning methods.
• Experimental verification confirms the algorithm's effectiveness in enhancing image reconstruction quality.