Problem
A PET-CT lab wants to shorten each scan from 90 seconds. Scanners are expensive, the radiopharmaceutical stays usable for only 10–12 hours, and an examination costs from 30 to 60 thousand rubles. A shorter scan raises tomograph throughput and spares the patient. But a shorter acquisition gives a noisier image, and the doctor must still be able to read it as well as the original. So the team needs a denoiser and a way to show that its output is diagnostically no worse.
Example from the article: Sber's medical-data laboratory set out to restore 60-second and 30-second scans to the quality of the standard 90-second scan.
Solution in brief
Treat it as supervised denoising. The training pairs come for free: some scanners export intermediate series with shorter acquisition times, so the noisy input and the clean 90-second target come from the same study. Train a network to predict the difference between noisy and clean images, and judge it with SSIM, RMSE and tumor-level SUV agreement.
Questions this case answers
- How to halve image acquisition time by restoring noisy images with a network trained on paired data that the device itself produces
- How to show that a restored image preserves quantitative measurements needed for decisions
- How to halve the scan time of a medical imaging study with a denoising neural network
- How to build a paired noisy/clean training set without extra patient exposure
Tools and methods
- ResNet, U-Net, CycleGAN, SwinIR — the four model families compared
- SSIM, RMSE — image-level quality metrics
- Bland-Altman plot on SUVmax and mean SUV — checks tumor measurements
- nnU-Net, cc3d — find tumor regions automatically
- Optuna — learning-rate search
Google or a one-time email code. No password.