ICANEWS

3D Rectified Flow for Efficient Ultra-Low-Dose Whole-Body PET Image Denoising

arXiv CS · · 3 min read · Engineering & Technology

Read research and analysis on 3D Rectified Flow for Efficient Ultra-Low-Dose Whole-Body PET Image Denoising published by ICANEWS, a global research journal for emerging researchers.

Key Takeaways

  • The proposed 3D Flow model reconstructs a full 3D volume in approximately 30 seconds, significantly faster than multi-hour inference for 3D DDPM baselines.
  • The model achieves favorable global image quality and lesion conspicuity compared to 3D DDPM and DDIM baselines, including on short-acquisition data.
  • Promising zero-shot transfer performance was observed across evaluated datasets and unseen dose levels (down to 1/100 of standard dose).

Why This Matters

Reducing radiation exposure in PET imaging is important for patient safety. This method addresses the noise challenges in ultra-low-dose imaging, potentially enabling safer diagnostic procedures while maintaining image quality and computational efficiency.

Overview

This research introduces a one-pass conditional 3D rectified flow (3D Flow) framework designed for denoising whole-body Positron Emission Tomography (PET) images, particularly those acquired at ultra-low doses. The framework integrates a novel optimized non-uniform sampling strategy and is trained with a one-pass linear-interpolant velocity-matching objective. The primary objective is to enhance diagnostic interpretation in ultra-low-dose PET imaging, which typically suffers from severe noise due to reduced radiation exposure.

Research Context

Patient safety concerns necessitate reducing radiation exposure in PET imaging. However, ultra-low-dose imaging protocols inherently result in increased image noise. This noise can adversely affect the diagnostic utility of the images if not appropriately mitigated. Existing 3D deep generative models, such as diffusion models, have demonstrated strong reconstruction fidelity in image enhancement tasks. Nevertheless, their practical application is often constrained by extended inference times. Conversely, faster 2D-based alternatives face challenges in maintaining volumetric consistency, a critical aspect for comprehensive whole-body PET imaging analysis.

Approach

The proposed method addresses the identified gap by developing a one-pass conditional 3D rectified flow framework. Key components of this approach include:

  • Conditional 3D Rectified Flow: The core of the framework is a 3D rectified flow model, which operates on the entire 3D volume.
  • Optimized Non-Uniform Sampling Strategy: A novel sampling strategy is incorporated to enhance the model's performance.
  • Training Objective: The model is trained using a one-pass linear-interpolant velocity-matching objective. This objective guides the model in learning the denoising process efficiently.

The implementation reconstructs a full 3D volume. Inference time measurements were performed to compare the proposed method with evaluated baselines.

Findings

The evaluation of the proposed 3D Flow framework yielded several key findings:

  • Computational Efficiency: The model demonstrated significantly faster inference times. It reconstructed a full 3D volume in approximately 30 seconds within the implemented environment. This contrasts sharply with the multi-hour inference times observed for the evaluated 3D Denoising Diffusion Probabilistic Model (DDPM) baseline.
  • Image Quality and Lesion Conspicuity: Evaluations, including zero-shot transfer to an independent clinical dataset, indicated that the proposed model achieved favorable global image quality and lesion conspicuity. This performance was observed in comparison to the evaluated 3D DDPM and Denoising Diffusion Implicit Models (DDIM) baselines. This favorable outcome was particularly noted even on challenging short-acquisition data.
  • Zero-Shot Transfer Performance: The method exhibited promising zero-shot transfer capabilities. This performance was observed across the evaluated datasets and extended to unseen dose levels, specifically down to 1/100 of the standard dose.
  • Artifact-Focused Visual Comparisons: While demonstrating strong performance, artifact-focused visual comparisons supported the need for further lesion-level validation.

Overall, the work presents a candidate approach that balances reconstruction fidelity with computational efficiency for ultra-low-dose whole-body PET image denoising.

Why This Matters

Reducing radiation exposure in PET imaging is important for patient safety. The development of an efficient denoising method for ultra-low-dose PET addresses the severe noise that can compromise diagnostic interpretation. By balancing reconstruction fidelity with computational efficiency, this approach offers a practical candidate for clinical implementation in ultra-low-dose whole-body PET imaging.

Potential Applications

This work presents a candidate approach for ultra-low-dose whole-body PET image denoising. This could facilitate clinical protocols that prioritize patient safety by minimizing radiation exposure without sacrificing image quality for diagnostic purposes.

Research Information

Institution
arXiv CS
Original Study
View Publication
Source
arXiv CS

About ICANEWS

ICANEWS is a global research journal for emerging researchers, publishing student and emerging researcher work across all fields.