Filter2Noise: a framework for interpretable and zero-shot low-dose CT image denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Image Denoising

Background:

• Deep learning excels at low-dose computed tomography (LDCT) denoising but lacks transparency, hindering radiologist trust.
• Existing zero-shot methods struggle with the spatially correlated noise inherent in CT imaging.
• Interpretability and verifiability are crucial for clinical adoption of AI-driven medical imaging tools.

Purpose of the Study:

• To develop a denoising method for LDCT that achieves deep learning performance while remaining fully interpretable and controllable by radiologists.
• To address the limitations of black-box models and zero-shot methods in handling CT's specific noise characteristics.
• To introduce a transparent, content-adaptive mathematical operator for medical image denoising.

Main Methods:

• Introduced Filter2Noise (F2N), an attention-guided bilateral filter adapting to local anatomy, replacing conventional deep networks.
• Developed Euclidean local shuffle to disrupt noise correlations and a multi-scale self-supervised loss for robust learning from single noisy images.
• Employed a lightweight attention module (3.6k parameters) to predict optimal filtering strategies based on tissue type, texture, and noise.

Main Results:

• F2N achieved 39.76 dB peak signal-to-noise ratio on the Mayo Clinic LDCT Grand Challenge, outperforming other zero-shot methods.
• The method used significantly fewer parameters (3.6k vs. 1.3M) compared to deep learning models.
• Clinical validation showed F2N elevated low-dose images to full-dose quality with no significant difference in contrast-to-noise ratio (p=0.10).

Conclusions:

• F2N successfully reconciles high performance with complete interpretability and user control in LDCT denoising.
• The visualizable filtering strategy allows radiologists to interactively refine denoising in critical regions.
• F2N provides a verifiable, retraining-free tool applicable across different scanners and protocols.