AI Cleaning Up CT Scans May Also Hide Tumors, Study Finds
Clearer medical images are useless if the AI smooths away early cancer signs.
Researchers tested AI that removes noise from low-dose CT scans. Two surprises: a single pass beats the usual multi-step refinement — because multi-step integration provably departs from the MMSE solution — and standard image scores can improve while low-contrast lesion detectability drops. Supervised denoisers gained about 4 dB PSNR, yet a channelized Hotelling observer revealed clinically relevant degradation that PSNR and SSIM missed. The key wasn't the flow machinery but the pairing: a one-step regressor trained on matched noisy pairs gave the best label-free result.
- More AI processing steps make denoised scans worse, not better — the opposite of what most researchers assumed.
- One popular method, Noise2Void, performed no better than doing nothing on CT scans because its assumptions about noise don't hold.
- AI-smoothed scans scored about 4 decibels 'better' on standard image tests, yet a simulated radiologist spotted fewer low-contrast lesions in them.
Why It Matters
Hospitals buying AI scan cleanup should verify it doesn't erase the early disease signs doctors need to find.