Abstract
Due to the effect of noise in brain MR images, it is difficult for the traditional fuzzy c-means (FCM) clustering algorithm to obtain desirable segmentation results. Combining the information of patch to reduce the effect of noise has been a focus of current research. However, the traditional patch model is isotropic, so that it would lose the structural information easily. In this paper, a novel fuzzy c-means model based on the spatial similarity information is proposed. To be anisotropy and preserve more structural information, this model takes both the non-local information and spatial structural similarity measurement (SSIM) between the image patches into consideration, and then a new distance function is established between every pixels and category centers for image segmentation. The efficiency of the proposed algorithm is demonstrated by extensive segmentation experiments using both synthetic and real brain MR images and by comparison with other state of the art algorithms.
Introduction: The Challenge of MRI Noise
Brain MRI segmentation—the process of partitioning an image into distinct tissue types such as Gray Matter (GM) and White Matter (WM)—is a cornerstone of modern neurological diagnosis. However, the inherent physics of MRI acquisition often introduces significant noise and artifacts. Traditional segmentation tools, most notably the Fuzzy C-means (FCM) clustering algorithm, are "spatial-blind." They classify pixels based solely on intensity, making them highly susceptible to misclassification in the presence of noise.
While previous attempts to incorporate spatial context (e.g., FCM_S, FCM_NLS) have improved noise robustness, they often rely on isotropic smoothing. This approach treats all directions equally, which effectively suppresses noise but at the cost of blurring sharp edges and destroying the slender, delicate anatomical structures characteristic of the human brain.
Methodology: Anisotropic Patch-Based Clustering
The proposed model addresses these limitations by introducing an anisotropic distance function that respects the directional nature of anatomical structures. Instead of clustering individual pixels, the algorithm operates on image patches, utilizing a dual-weighting scheme:
- Non-local (NL) Information: The model searches a local window for patches with similar neighborhood configurations, allowing it to distinguish between actual tissue structures and random noise.
- Structural Similarity (SSIM): By integrating SSIM—a metric typically used for image quality assessment—directly into the clustering process, the algorithm can compare luminance, contrast, and structure. This ensures that only structurally similar regions influence the classification of a given pixel.
By adaptively balancing these weights using a parameter ($\lambda$), the model maintains high segmentation accuracy even as noise levels fluctuate.
Results and Discussion
The researchers evaluated the proposed method against standard FCM and several state-of-the-art variants (FCM_S1, FCM_S2, and FCM_NLS) using both synthetic and real-world clinical scans.
Quantitative Performance
Using the Jaccard Similarity (JS) index as a benchmark, the proposed method demonstrated superior accuracy across all noise levels. As shown in the table below, the performance gap becomes particularly pronounced at higher noise levels.
At a 7% noise level, the proposed method achieved a JS of 89.49% for White Matter, significantly outperforming the standard FCM (69.17%) and the previous best-performing variant, FCM_NLS (84.78%).
Qualitative Analysis
Visual comparisons further highlight the model's ability to preserve fine details. While isotropic methods tend to produce "blobby" segmentations where thin tissue "fingers" are lost, the proposed anisotropic model keeps boundaries sharp and preserves slender structures.
Conclusion
By shifting the focus from pixel intensity to structural context via SSIM-guided patch clustering, this novel FCM model provides a robust solution for brain MRI segmentation. It effectively bridges the gap between noise suppression and structural preservation, offering a more reliable tool for clinical neuroimaging analysis. Future work may explore the application of this framework to other noisy medical imaging modalities, such as ultrasound or CT.
Summary updated