Noise Reduction in CT Using Learned Wavelet-Frame Shrinkage Networks.

Area of Science:

• Medical Imaging
• Computer Vision
• Signal Processing

Background:

• Encoding-decoding (ED) Convolutional Neural Networks (CNNs) excel at noise reduction.
• The theory of deep convolutional framelets (TDCF) links CNNs to signal processing, showing ReLU CNNs often induce low-rankness and lack perfect reconstruction (PR).

Purpose of the Study:

• To explore CNNs that satisfy PR conditions for noise reduction.
• To propose novel CNN architectures for enhanced denoising in medical imaging.

Main Methods:

• Investigated CNNs meeting PR conditions, demonstrating soft shrinkage and PR are achievable.
• Proposed the learned wavelet-frame shrinkage network (LWFSN) and its residual version (rLWFSN).
• LWFSN's ED path adheres to PR conditions, with shrinkage based on linear threshold expansion.

Main Results:

• The proposed LWFSN and rLWFSN comply with PR conditions.
• LWFSN utilizes <1% of parameters compared to conventional CNNs, offering fast inference and low memory usage.
• Achieved performance comparable to state-of-the-art methods like TF U-Net and FBPConvNet in low-dose CT denoising.

Conclusions:

• CNNs meeting PR conditions enable effective soft shrinkage and denoising.
• LWFSN and rLWFSN offer a computationally efficient alternative for low-dose CT denoising with high performance.