Boost the adversarial learning with Fourier regulator : bias-field correction on MRI

Abstract

In magnetic resonance imaging, signal intensity inhomogeneities due to intrinsic bias field pose a significant challenge for automated medical image analysis. Conventional methods to mitigate these effects, such as N4ITK, are time-consuming and unstable. The exploration of deep learning alternative approaches is still at an unknown stage. Previous studies have obtained preliminary results in GAN-based models, but we found the difficulty in aligning bias-corrected image domains with clean image domains during adversarial learning may affect the retention of normal organizational structures. Therefore, we propose a novel Fourier regulator structure that can be integrated into the general adversarial learning framework. It explicitly decouples different levels of semantic features based on the Fourier field and utilizes explicit feature learning to enhance intrinsic coherence and promote more organized domain alignment. By separating amplitude and phase features as well as splitting low and high-frequency information, our model preserves organizational details more efficiently and explicitly separates intensities across organizational boundaries. During the training process of adversarial learning, the generator generates the target domain while the regulator and discriminator are fixed; the regulator and discriminator are updated in parallel while the generator is fixed. Such a learning approach extends the original min-max optimization problem of adversarial learning to a multi-player mix-max optimization problem. The discriminator can quickly draw the generative domain closer to the target domain, while the regulator aligns the distance to the target domain in a more explicit feature-learning manner. Evaluated on the OASIS and BrainWeb datasets, our model outperforms traditional and deep learning methods to enhance homogeneity. It also shows consistent performance in other image reconstruction tasks, demonstrating its generalization capabilities.