Wavelet subband discriminator for efficient unsupervised chest X-ray image restoration

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Chest X-ray (CXR) imaging is crucial for non-invasive diagnosis.
• Image quality is vital for accurate clinical diagnosis but often degraded by noise and scatter.
• Existing methods like WavCycleGAN face limitations in processing diverse artifacts and require complex, multi-step procedures.

Purpose of the Study:

• To develop a novel unsupervised scheme for restoring chest X-ray (CXR) images.
• To overcome limitations of existing methods by simplifying the restoration process and improving artifact removal.
• To achieve comparable or superior artifact removal performance to WavCycleGAN without additional testing phase processing.

Main Methods:

• Introduced a novel wavelet subband discriminator integrated with CycleGAN or switchable CycleGAN.
• Applied wavelet transform only during the training phase for discriminators.
• Utilized an image-domain cycle-consistency loss to prevent artifacts from the wavelet domain.
• Employed frequency-specific wavelet subband discriminators for comprehensive artifact removal across all subbands.

Main Results:

• Demonstrated competitive performance in noise and scatter removal for CXRs compared to existing methods.
• Confirmed no additional processing steps are needed in the test phase.
• Showcased the flexibility of the wavelet subband discriminator with switchable CycleGAN for adjustable artifact removal levels.

Conclusions:

• The proposed wavelet subband discriminator offers an efficient unsupervised approach for CXR image reconstruction when combined with CycleGAN or switchable CycleGAN.
• This method eliminates the need for additional testing phase processing and avoids unnatural artifacts common in wavelet domain image restoration.
• The developed technique holds potential for broad application in various chest X-ray imaging scenarios.