Abstract
Medical imaging, particularly Magnetic Resonance Imaging (MRI), frequently suffers from noise that compromises diagnostic accuracy. Traditional denoising filters often blur critical edges, while modern deep learning approaches require vast amounts of labeled data. This paper introduces a "denoising via repainting" framework that reconstructs images through progressive Bézier-path redrawing. By combining anisotropic filtering with multi-scale vectorization, the proposed method achieves state-of-the-art performance across diverse MRI datasets without the need for large-scale training.
Introduction
Magnetic Resonance Imaging (MRI) is essential for visualizing fine anatomical structures, yet it is often plagued by noise. Existing solutions generally fall into two categories: conventional filters and deep learning models. Conventional methods, such as wavelet thresholding, tend to over-smooth textures and edges. Conversely, deep learning models like DnCNN are highly effective but require massive, clean datasets and often struggle to generalize across different noise types or imaging modalities.
The "denoising via repainting" approach offers a data-efficient alternative. It treats image reconstruction as a geometric optimization problem, ensuring structural integrity and edge preservation by representing image components as vector shapes.
Methods
The proposed framework operates through a multi-stage pipeline:
- Scale-Space Pyramid Construction: The algorithm generates a sequence of images using anisotropic Gaussian blur. This "smart" blur reduces noise while respecting anatomical edges, creating a pyramid ranging from a heavily smoothed base ($I_0$) to a detailed version ($I_T$).
- Progressive Component Redrawing: Starting at the coarsest level, the system segments the image into coherent regions. These regions are "repainted" using parametric cubic Bézier paths. This vector-based representation ensures clean, continuous boundaries that are inherently resistant to pixel-level noise.
- Coarse-to-Fine Optimization: As the process moves to finer scales, the algorithm identifies and adds new Bézier paths to capture intricate details not visible at coarser levels.
- Loss Functions: Accuracy is maintained using Mean Square Error (MSE) and a specialized "Xing-loss," which prevents topological artifacts like self-intersecting curves.
Results and Discussion
The method was evaluated against supervised models (DnCNN, Restormer) and unsupervised techniques (Anisotropic filtering, Beltrami) across several datasets, including AxFLAIR (brain), Cor-PD (knee), and FastMRI (breast and prostate).
Quantitative Performance
As shown in the experimental results, the "Ours" method consistently outperformed both trained and untrained baselines in terms of Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). For instance, on AxFLAIR scans with $\sigma=10$, the method achieved a PSNR of 39.30, significantly higher than DnCNN's 38.10.
Qualitative Superiority
Visual inspections reveal that while deep learning models may produce mottled artifacts or over-smoothed textures, the repainting method maintains sharp, clean boundaries that closely align with the ground truth. By treating images as geometry rather than just pixel grids, the framework preserves shape coherence more effectively than pixel-based methods.
Conclusion
Denoising via repainting provides a robust, zero-shot solution for medical image enhancement. While the iterative optimization process is more computationally intensive than standard neural network inference, its ability to generalize without training data makes it highly versatile. Future work will focus on extending the framework to 3D volumetric data and automating hyperparameter selection through meta-learning.
Summary updated