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

Deep learning has achieved remarkable progress in low-dose computed tomography (LDCT) denoising; however, radiologists struggle to trust black-box models they cannot verify or control. Zero-shot methods eliminate training data requirements but fail on computed tomography's (CT) spatially correlated noise. We demonstrate that a transparent mathematical operator, when made content-adaptive, can match deep learning performance while remaining fully interpretable.

We introduce Filter2Noise (F2N), which replaces conventional deep networks with an attention-guided bilateral filter that adapts to local anatomy. A lightweight attention module (3.6k parameters) predicts optimal filtering strategies for each image region by analyzing tissue type, texture, and noise characteristics. To enable robust learning from a single noisy image with correlated noise, we develop Euclidean local shuffle, which strategically disrupts noise correlations while preserving anatomical structure, and a multi-scale self-supervised loss that enforces consistency across resolutions.

On the Mayo Clinic LDCT Grand Challenge, F2N achieves 39.76 dB peak signal-to-noise ratio, outperforming the next-best zero-shot method by 1.88 dB, while using 360× fewer parameters (3.6k versus 1.3M). Clinical validation on photon-counting CT demonstrates that F2N elevates low-dose images to full-dose quality (no statistically significant difference in contrast-to-noise ratio, $p=0.10$ ). The learned filtering strategy is fully visualizable: parameter maps reveal content-aware behavior. Radiologists can interactively adjust these parameters post-training to refine denoising in diagnostically critical regions.

F2N reconciles competitive performance with complete interpretability and user control, providing radiologists with a verifiable tool that works across scanners and protocols without retraining.