Noise2Flow deep dive: denoising a single medical image by trusting the direction, not the destination
maier_ak · x · 2026-11-16
Andreas Maier's Substack deep dive on Noise2Flow (Journal of Medical Imaging, led by Yipeng Sun, with HIT and Peking Union Medical College Hospital).
- Problem: medical imaging lacks training data—clean CTs need higher radiation dose, live cells can't be imaged twice, cross-scanner generalization is poor. Existing single-image methods either assume independent pixel noise (false for reconstructed low-dose CT and real microscopy) or, like Filter2Noise/Noise2Detail, leave output visibly noisy.
- Core idea: an imperfect denoiser is often right about direction and wrong about distance—the step from noisy image to output usually points the right way (verified against ground truth). The method learns that direction from one image, fits a step length, and walks 15% past the imperfect result, with one network pass at the end.
- Results: 34.91 dB on fluorescence microscopy (+0.52 dB over Filter2Noise), 28.82 dB on low-dose CT (+0.91 dB), better for 9 of 10 patients; gains are moderate, pixel-metrics only, no reader study yet.
Related event: Noise2Flow Denoises Single Noisy Medical Images Without Training Data(3 posts)→
More from Research
- Slow models are a disaster in robotics: researcher dives into fixing latency for large VLAs — kvablack · 2026-11-16
- ICL alone may enable substantial continual learning, plot shows agents improve by Month 3 — xennygrimmato_ · 2026-11-16
- 1% erbium doping wrecks reactive ion etching of TeO₂ waveguides — jwt0625 · 2026-11-16
- Noise2Flow: trust the denoiser's direction, then walk 15% past its output — maier_ak · 2026-11-16
- Noise2Flow targets the hardest denoising case: one noisy image, no ground truth — maier_ak · 2026-11-16
- VLM memory survey from 8 institutions: 4D taxonomy covering 300+ papers and ~70 benchmarks — 机器之心 · 2026-11-16