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

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.

💡 Research Summary

The paper introduces a novel zero‑shot deep learning framework for magnetic resonance imaging (MRI) bias field correction that requires no pre‑training data. Traditional bias correction methods fall into two categories: prospective approaches that modify the scanner hardware or acquisition protocol, and retrospective approaches that process the acquired images. Among retrospective methods, the most widely used non‑learning technique is N4ITK, which iteratively estimates a smooth bias field by histogram‑based optimization. While effective, N4 is computationally intensive and can produce unrealistic results when the bias variation is large. Recent supervised and unsupervised deep learning methods have shown promising performance, but they rely on large annotated datasets or simulated bias fields, limiting their generalizability and increasing the cost of data collection.

The proposed method reframes bias correction as a zero‑shot problem: given a single corrupted volume, the algorithm learns a per‑voxel correction parameter α and an estimated bias map ˆB directly at test time. The core of the system is a lightweight 3‑D convolutional neural network (CNN) built from depthwise separable convolutions (DSC). The network processes an 8× down‑sampled version of the input volume, resulting in roughly 3 000 trainable parameters—orders of magnitude fewer than typical CNNs. This compact design dramatically reduces memory consumption and enables rapid online optimization.

Two pixel‑wise maps are produced by the network: (1) α, which drives an iterative homogeneity correction function HC defined as
 HCₙ = HCₙ₋₁ + α·HCₙ₋₁·(1 – HCₙ₋₁)
with four iterations (n = 4) empirically chosen to balance accuracy and speed; and (2) ˆB, a bias field estimate constrained to the