Weak edge detection using Mean-shift filtering and histogram enhancement

Feng Ji, Xinbo Gao, Songyun XieJournal of Image and Graphics 2012 · DOI

Abstract: Accurate localization of functional regions in the brain requires precise extraction of biological tissues, such as the skull, from functional Magnetic Resonance Imaging (fMRI) data. However, existing limitations in MRI technology often result in "weak edges" for these structures, making them difficult to isolate using standard techniques. This paper proposes an effective method for weak edge extraction. First, Mean-shift filtering is employed to suppress noise without losing critical target information. Subsequently, a targeted histogram enhancement is applied to the skull region based on local grayscale distribution characteristics. Finally, standard edge detection algorithms are used to extract the skull boundary. Experimental results demonstrate that this approach effectively isolates weak skull edges in fMRI images, overcoming the limitations of traditional filters.

Introduction

In neuroimaging, the precise mapping of functional brain areas depends on an accurate anatomical context, primarily defined by the skull boundary. However, fMRI data is inherently noisy, and the signal intensity of bone tissue is typically low, often blending into background interference. Standard edge detection operators, such as Sobel or Canny, frequently fail in these contexts, producing fragmented edges or misidentifying noise as biological structure. This research addresses the need for a robust pre-processing pipeline capable of suppressing noise while preserving and amplifying these subtle "weak edges."

Methodology

The proposed workflow consists of three integrated stages designed to enhance signal-to-noise ratios specifically for low-intensity anatomical structures.

1. Mean-shift Filtering

To combat noise without the blurring artifacts typical of Gaussian filters, the authors utilize Mean-shift filtering. As a non-parametric clustering technique, Mean-shift shifts each pixel toward the local mode (the densest point) of its feature space. This process smooths internal regions while maintaining sharp transitions at boundaries, effectively "cleaning" the image while preserving the underlying geometry of the skull.

2. Targeted Histogram Enhancement

Post-filtering, the skull tissue often resides within a narrow, low-intensity grayscale range—typically between 25 and 45 in the studied dataset. The authors apply a piecewise non-linear transformation to "stretch" this specific interval. By amplifying this range by a factor of three while suppressing the surrounding background and internal brain tissue, the skull's signal is significantly boosted relative to its environment.

3. Multi-slice Accumulation

To further distinguish reliable biological edges from transient artifacts, the method incorporates a temporal/spatial consistency check. By processing multiple consecutive fMRI slices, the system implements an accumulation module: a pixel is only classified as a definitive "edge" if it consistently appears in the same spatial location across more than 80% of the processed slices.

Experimental Setup and Datasets

The methodology was validated using a dataset provided by the Fourth Military Medical University, consisting of 12 layers of fMRI scans totaling 600 images. Each image has a resolution of 128 x 128 pixels. The scans were captured under controlled conditions with the subject's head fixed, ensuring that the multi-slice accumulation could be performed without the need for complex image registration.

Discussion and Key Insights

A critical observation made during the study was the "double-edged sword" nature of filtering. While Mean-shift is superior at noise suppression, it can also marginally reduce the amplitude of already faint signals. This insight necessitates the subsequent histogram enhancement step; filtering alone is insufficient for weak edge detection.

Furthermore, the study highlights the efficacy of domain-specific boosting. Unlike global histogram equalization, which can introduce artifacts and wash out details, the "surgical" enhancement of a specific grayscale signature allows for the isolation of the skull without distorting the rest of the image. By leveraging the 3D nature of medical data through the 80% voting rule, the researchers significantly reduced false positives caused by random noise.

Results and Conclusion

Qualitative analysis shows that the proposed pipeline successfully reveals the skull boundary, which is nearly invisible in raw fMRI data and ignored by standard Canny operators. The Mean-shift step effectively removes "salt and pepper" noise, while the targeted enhancement allows the Canny operator to produce a nearly continuous, closed-loop boundary.

The final processed images permit the creation of a clean binary mask (using a threshold of 120), effectively separating the skull from internal brain structures. While the specific grayscale parameters may require recalibration for different scanners, this framework provides a powerful tool for improving the accuracy of anatomical mapping in functional neuroimaging. Future work will focus on automating the identification of target grayscale ranges and exploring adaptive kernels for the Mean-shift process.

Summary updated

BibTeX
@article{ji2012weak,
  author = {Feng Ji and Xinbo Gao and Songyun Xie},
  title = {{Weak edge detection using Mean-shift filtering and histogram enhancement}},
  journal = {Journal of Image and Graphics},
  year = {2012},
  doi = {10.11834/jig.20120606},
}
Figure