U-TTT: Towards Generalizable PET Image Denoising via Test-Time Training

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Science & Research — Health & Medical Research, Mathematics & Computational Sciences · Depth: Expert, long

Summary

U-TTT is a novel U-shaped deep learning model designed for robust Positron Emission Tomography (PET) image denoising, specifically addressing performance degradation under distribution shifts. It integrates Test-Time Training (TTT) layers to dynamically adjust model parameters during inference, adapting to individual test data characteristics like varying dose levels or scanner types. A key innovation is its dual-domain adaptation mechanism, featuring a Spatial Test-Time Training (S-TTT) layer for correcting spatial structural degradations and a Frequency Test-Time Training (F-TTT) layer for suppressing global noise spectra and restoring high-frequency details. This approach allows U-TTT to achieve state-of-the-art PET denoising performance, outperforming methods like VQPET by an average of 0.80 dB in PSNR and 0.0028 in SSIM, and reducing lesion error by 0.0154. Experiments confirm its superior generalization across unseen dose levels and scanners.

Key takeaway

For Machine Learning Engineers developing PET image denoising solutions, you should consider integrating Test-Time Training (TTT) layers into your models. This approach dynamically adapts to unseen scanner types and dose levels, significantly improving generalization. Implement dual-domain adaptation with spatial and frequency TTT layers to comprehensively address noise and structural degradation. This ensures robust clinical deployment and superior performance compared to static models.

Key insights

U-TTT uses dual-domain Test-Time Training layers for robust PET image denoising under distribution shifts.

Principles

• Fixed-parameter models degrade under distribution shifts.
• Test-Time Training layers enable dynamic adaptation.
• Dual-domain adaptation improves noise suppression.

Method

U-TTT employs a U-shaped encoder-decoder with S-TTT and F-TTT blocks. These layers perform feature reconstruction tasks, dynamically updating inner model parameters via gradient descent during inference.

In practice

• Integrate TTT layers into U-Net architectures for adaptation.
• Design inner models for 3D vision tasks, not just 1D.
• Combine spatial and frequency domain adaptation for comprehensive denoising.

Topics

• PET Image Denoising
• Test-Time Training
• Distribution Shift
• U-Net Architecture
• Medical Imaging
• Deep Learning Generalization

Code references

• Yaziwel/U-TTT
• Yaziwel/U-TTT

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• A Decision-Theoretic View of Test-Time Training: When, How Far, and Which Directions to Adapt — TTT's effectiveness hinges on spectrally matching update steps and aligning update directions with query relevance.
• Dual-Domain Equivariant Generative Adversarial Network for Multimodal CT-PET Synthesis — Combining dual-domain learning with geometric equivariance significantly enhances multimodal medical image synthesis accuracy and robustness.
• ROBUST-WT: Robust Uncertainty-aware Segmentation Transform via Whitening and Training Enhancements — ROBUST-WT demonstrates that targeted training enhancements can substantially improve medical image segmentation robustness without altering core architecture.
• SHIFT: Robust Double Machine Learning for Average Dose-Response Functions under Heavy-Tailed Contamination — SHIFT robustifies DML for ADRF by using post-GNC residual MAD for outlier detection, improving accuracy and providing an outlier mask.
• Fast Spatial Memory with Elastic Test-Time Training — Elastic Test-Time Training stabilizes fast-weight updates in 4D reconstruction, enabling robust multi-chunk adaptation for long sequences.
• Test Time Training for Supervised Causal Learning — TTT-SCL improves Supervised Causal Learning's real-world applicability by dynamically generating test-aligned training sets.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.