Abstract
Deep learning models for medical image artifact reduction typically require large, paired datasets of clean and noisy images, which are often unavailable in clinical practice. This paper introduces OSAR (One-Shot medical image Artifact Reduction), a framework that synthesizes training data directly from a single patient scan at test-time. By isolating real-world artifacts from uniform backgrounds and training a specialized attentive network on-the-fly, OSAR achieves significant improvements in signal-to-noise ratio (SNR) and visual fidelity without the need for external training data.
Introduction
The efficacy of deep learning in medical imaging is often hampered by the "data hunger" of artifact reduction models. Traditional supervised approaches rely on thousands of paired images—identical scans with and without artifacts. Since true "clean" ground-truth images are rarely available clinically, researchers often resort to simulated noise. However, these simulations frequently fail to capture the complexity of real-world clinical artifacts, such as patient motion or machine interference, leading to poor performance on actual patient data. Existing "zero-shot" methods attempt to circumvent this but often struggle when artifact intensity overlaps with tissue intensity.
Methods
OSAR addresses these challenges through a "One-Shot" approach, where the model is trained and tested on the same specific input image. The workflow is divided into two primary stages:
- Internal-Data-Synthesis-Network (IDSN): A radiologist identifies a few regions of interest (ROIs) where artifacts appear against a uniform background (e.g., air or fat). A lightweight CNN classifier learns to distinguish these artifact patterns from actual tissue. The system then superimposes these isolated artifacts onto other parts of the image, creating a large, patient-specific synthetic dataset of paired "dirty" and "clean" patches.
- Attentive-Artifact-Reduction-Network (AARN): This network serves as the restoration engine. It employs a recurrent attention mechanism to pinpoint artifact locations and an autoencoder to focus its restorative capacity on those regions while preserving anatomical details. The network's compact design allows for rapid convergence, typically within 2–3 epochs, making "on-the-fly" training feasible in a clinical setting.
Key Results and Insights
The framework was evaluated using clinical CT (cardiac) and MRI data, covering various artifact types including Poisson noise, motion artifacts, Gibbs ringing, and aliasing.
- Superior Performance: OSAR demonstrated a 103.6% overall improvement in Signal-to-Noise Ratio (SNR). In CT scans specifically, it outperformed competing methods like ZSAR and CCADN by a factor of 2 to 3.
- Visual Fidelity: Qualitatively, OSAR effectively reduced "striping" artifacts without the "waxy" over-smoothing typical of general-purpose denoisers trained on simulated data.
- The "Uniform Background" Trick: By isolating noise where it exists alone (e.g., in the air surrounding a patient), the model avoids the bias of simulated noise and learns the exact distribution present in the current scan.
Limitations and Future Work
While the method currently requires minimal manual intervention (annotating as few as 7 ROIs), future work aims to automate this process to move toward a fully "Zero-Shot" system. Additionally, OSAR is currently optimized for additive artifacts; further research is needed to address transformative artifacts that fundamentally warp underlying tissue geometry.
Conclusion
OSAR represents a significant step toward practical, patient-specific artifact reduction. By leveraging the information within a single image, it bypasses the need for massive external datasets and provides superior performance on real-world clinical noise, offering a faster and more accurate solution for diagnostic imaging.
Summary updated