Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Zero-shot Bias Correction: Efficient MR Image Inhomogeneity Reduction Without Any Data

Hongxu Yang, Edina Timko, Brice Fernandez

TL;DR

This work tackles MRI inhomogeneity caused by a bias field and additive noise by introducing a zero-shot, data-free deep learning approach that requires no pre-training. It decomposes bias correction into online homogeneity refinement guided by per-voxel parameters and a predicted bias field, enforced through an image-prior constraint and a lightweight CNN with approximately $3\text{k}$ parameters. The method optimizes $L(\theta)$ at test time using $L_{hc}$ (including smoothing, spatial, and exposure terms) and $L_{prior}$ (alignment to the observed image via $|Y - \hat{X}\hat{B}|_{1}$), yielding fast, stable convergence. Across five MRI datasets, the approach matches or outperforms the data-free N4 baseline in accuracy while delivering substantial speedups (about $3$ seconds on an NVIDIA $A100$), enabling faster preprocessing in automated MRI pipelines.

Abstract

In recent years, deep neural networks for image inhomogeneity reduction have shown promising results. However, current methods with (un)supervised solutions require preparing a training dataset, which is expensive and laborious for data collection. In this work, we demonstrate a novel zero-shot deep neural networks, which requires no data for pre-training and dedicated assumption of the bias field. The designed light-weight CNN enables an efficient zero-shot adaptation for bias-corrupted image correction. Our method provides a novel solution to mitigate the biased corrupted image as iterative homogeneity refinement, which therefore ensures the considered issue can be solved easier with stable convergence of zero-shot optimization. Extensive comparison on different datasets show that the proposed method performs better than current data-free N4 methods in both efficiency and accuracy.

Zero-shot Bias Correction: Efficient MR Image Inhomogeneity Reduction Without Any Data

TL;DR

This work tackles MRI inhomogeneity caused by a bias field and additive noise by introducing a zero-shot, data-free deep learning approach that requires no pre-training. It decomposes bias correction into online homogeneity refinement guided by per-voxel parameters and a predicted bias field, enforced through an image-prior constraint and a lightweight CNN with approximately parameters. The method optimizes at test time using (including smoothing, spatial, and exposure terms) and (alignment to the observed image via ), yielding fast, stable convergence. Across five MRI datasets, the approach matches or outperforms the data-free N4 baseline in accuracy while delivering substantial speedups (about seconds on an NVIDIA ), enabling faster preprocessing in automated MRI pipelines.

Abstract

In recent years, deep neural networks for image inhomogeneity reduction have shown promising results. However, current methods with (un)supervised solutions require preparing a training dataset, which is expensive and laborious for data collection. In this work, we demonstrate a novel zero-shot deep neural networks, which requires no data for pre-training and dedicated assumption of the bias field. The designed light-weight CNN enables an efficient zero-shot adaptation for bias-corrupted image correction. Our method provides a novel solution to mitigate the biased corrupted image as iterative homogeneity refinement, which therefore ensures the considered issue can be solved easier with stable convergence of zero-shot optimization. Extensive comparison on different datasets show that the proposed method performs better than current data-free N4 methods in both efficiency and accuracy.
Paper Structure (11 sections, 8 equations, 2 figures, 1 table)

This paper contains 11 sections, 8 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Method
  3. Methodology
  4. CNN Model Design
  5. Online Optimization
  6. Experiments and Results
  7. Dataset
  8. Experiment settings
  9. Experiment results
  10. Conclusions
  11. Acknowledgement

Figures (2)

  • Figure 1: Top: The light-weight zero-shot inhomogeneity correction model. The input 3D volumetric data is processed by feature extraction blocks to generate parametric maps for iterative homogeneity refinement and image prior loss. DSC: Depthwise Separable Convolutions. The proposed model has around 3k trainable parameters. Bottom: Iterative inhomogeneity correction with the predicted $\alpha$ map per pixel.
  • Figure 2: Example images of comparisons. From top to bottom: IXI, ADNI, OASIS, Dixon-water and Dixon-fat. First column: original image, second column: N4 result, third column: the results of the zero-shot bias correction.