Enabling self-supervised learned primal dual with Noise2Inverse

2026-06-25 · Source: Takara TLDR - Daily AI Papers · Field: Science & Research — Mathematics & Computational Sciences, Engineering & Applied Sciences, Health & Medical Research · Depth: Expert, medium

Summary

Antti Sällinen, Siiri Rautio, Santeri Kaupinmäki, and Andreas Hauptmann introduce Noise2Inverse Learned Primal-Dual (N2I-LPD), a self-supervised method for X-ray computed tomography (CT) reconstruction. This approach addresses the ill-posed inverse problem in low-dose and sparse-angle CT settings. In these scenarios, ground-truth data for supervised training is often unavailable. N2I-LPD extends the Noise2Inverse framework to the Learned Primal-Dual algorithm. It trains an iterative reconstruction operator without ground-truth images. This is achieved by exploiting the statistical independence of noise in distinct measurements, specifically regarding the CT-scan's angular rotation. Comparative results show N2I-LPD achieves improved reconstruction quality. It outperforms classical methods and U-Net based neural networks trained within the Noise2Inverse framework. This demonstrates its potential for practical CT imaging scenarios lacking ground-truth data.

Key takeaway

For Machine Learning Engineers developing CT reconstruction algorithms, especially when ground-truth data is scarce, N2I-LPD offers a robust solution. You can achieve improved reconstruction quality by leveraging its self-supervised training, which exploits noise independence across CT-scan angular rotations. This eliminates the need for expensive labeled datasets. Consider integrating N2I-LPD or similar self-supervised primal-dual approaches into your workflow to enhance practical CT imaging applications.

Key insights

N2I-LPD enables self-supervised CT reconstruction by extending Noise2Inverse to Learned Primal-Dual, using noise independence for training without ground truth.

Principles

• CT reconstruction is an ill-posed inverse problem.
• Supervised learning needs ground-truth data.
• Noise independence enables self-supervised training.

Method

N2I-LPD extends Noise2Inverse to Learned Primal-Dual. It trains an iterative reconstruction operator by exploiting statistical independence of noise in distinct CT measurements, specifically regarding angular rotation, without ground-truth images.

In practice

• Apply N2I-LPD in low-dose CT imaging.
• Use N2I-LPD when ground-truth data is scarce.
• Combine learned operators with self-supervised strategies.

Topics

• X-ray Computed Tomography
• Image Reconstruction
• Self-supervised Learning
• Learned Primal-Dual
• Noise2Inverse
• Inverse Problems

Code references

• facebookresearch/vissl

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

Related on AIssential

• CRIS: Cross-Plane Self-Supervised Isotropic Restoration for Anisotropic Volumetric Imaging Across Modalities — CRIS enables isotropic volumetric image restoration using a self-supervised, cross-plane approach without paired ground truth.
• Learned Non-Maximum Suppression for 3D Object Detection — Learned filtering modules, D2D-Rescore and GossipNet3D, enhance 3D object detection by replacing heuristic NMS with detection-to-detection relation leveraging.
• Fusing Transferred Priors and Physics-based Decomposition for Underwater Image Enhancement — A novel UIE method fuses transferred priors and physics-based decomposition to enhance underwater images without paired labels.
• Plug-and-Play image restoration with Stochastic deNOising REgularization — SNORE improves PnP image restoration by applying denoisers to appropriately noisy images, enhancing performance and theoretical guarantees.
• 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.
• Limits of spectral learning under noise — Spectral learning exhibits a fundamental noise threshold that limits functional structure recovery from noisy data.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.