Abstract
Magnetic Resonance (MR) imaging is essential for disease diagnosis, yet it frequently suffers from intensity inhomogeneity, or "bias fields," caused by hardware limitations. These artifacts impede both qualitative and quantitative medical analysis. While unsupervised deep learning models have been proposed to address this, they often neglect image structure and bias field smoothness, leading to distorted results. This paper introduces S2DNets, a self-supervised dual-network framework that incorporates piece-wise structural constraints and bias field smoothness. By training two sub-networks to learn from each other without clean reference data, S2DNets effectively removes non-uniform intensity while retaining critical anatomical details. Extensive experiments on clinical and simulated datasets demonstrate that S2DNets outperforms conventional and state-of-the-art deep learning models.
Introduction
Intensity inhomogeneity manifests as smooth, low-frequency brightness variations across an MRI scan. While human observers can often overlook these artifacts, automated algorithms—such as those used for brain volume measurement or lesion detection—are easily misled. A darker patch of white matter caused by a bias field might be misclassified as gray matter, compromising diagnostic accuracy. Existing deep learning solutions typically require rare "clean" ground-truth data or fail to preserve sharp anatomical edges during the smoothing process. S2DNets addresses these challenges by treating structure preservation and bias correction as coupled tasks.
Methods: The S2DNets Architecture
S2DNets utilizes a dual-network architecture designed to address the limitations of global appearance learning. The framework operates on two primary fronts:
- The Clustering Part: This sub-network focuses on image structure. It leverages the assumption that a clean MRI is "piece-wise constant," meaning specific tissue types (e.g., gray vs. white matter) should have uniform intensities. Fuzzy clustering is used to generate structural maps, effectively narrowing the intensity distribution and excluding the bias field.
- The Bias Estimation Part: This sub-network estimates the artifact itself. Because bias fields are naturally smooth, the network is constrained to produce low-frequency outputs.
Unified Constraints
The core innovation of S2DNets lies in its unified objective function. It combines structural constraints (ensuring sharp edges) with smoothness constraints (using Gaussian kernels and Total Variation (TV) loss). This dual-learning approach is entirely self-supervised; the networks learn from each other to find a version of the image that resembles real tissue and a noise component that resembles a real bias field.
Experimental Results
The researchers evaluated S2DNets using the Human Connectome Project (HCP) clinical dataset (30,000 T1-weighted slices) and the BrainWeb simulated dataset.
Quantitative Performance
S2DNets achieved state-of-the-art results across multiple metrics. On the HCP dataset, it reached a PSNR of 28.40 and an SSIM of 0.979, significantly outperforming industry standards like N4 and recent deep learning models like IRNet. Furthermore, S2DNets achieved the lowest Coefficient of Variation (CV), indicating superior intensity uniformity within tissue types.
Qualitative Comparison
Visual analysis confirms that S2DNets preserves anatomical details that other models often blur. While methods like RCGAN or IRNet may leave "ghost-like" traces of anatomy in the estimated bias field, S2DNets maintains a smooth field that does not "steal" structural information.
Downstream Impact
To validate clinical utility, corrected images were processed through segmentation tools (U-Net and FSL). S2DNets-corrected images yielded the highest Dice scores, proving that superior bias correction directly enhances the accuracy of automated medical analysis.
Conclusion
S2DNets provides a robust, self-supervised solution for MR bias field correction. By integrating structural and smoothness constraints into a dual-network framework, it overcomes the data scarcity and distortion issues prevalent in existing deep learning methods. Future work will explore extending this 2D architecture to 3D volumes and applying the dual-network principle to other medical imaging modalities.
Summary updated