Radon-Like features and their application to connectomics

Ritwik Kumar, Amelio Vázquez Reina, Hanspeter PfisterWorkshops @ CVPR 2010 · DOI

Introduction

The field of Connectomics seeks to map the intricate wiring of the brain at nanometer resolution using Electron Microscopy (EM). However, the sheer volume of data—millions of neurons in a single mouse brain—precludes manual annotation. Automated segmentation is hindered by the noisy nature of EM images and the complex textures of biological structures like mitochondria and vesicles. This paper proposes Radon-Like Features (RLF) as a robust, computationally efficient descriptor tailored for these challenges.

Methodology

RLFs are inspired by the Radon Transform but adapted for local structural context. The process involves scanning the image along lines at multiple orientations to build a distribution of features for every pixel.

  1. Knots: Salient points, typically derived from an edge map, serve as "knots" along the scanning line.
  2. Extraction Functions: Between any two knots, information is aggregated using specific functions such as mean intensity, minimum value, or the distance between knots. This allows the framework to be tuned for different biological structures.
  3. Directional Sampling: By scanning from multiple angles (e.g., every 2 degrees), the system captures the full geometric context of the tissue.
Figure 6: Conceptual diagram of the RLF framework, where information is aggregated between structural 'knots' along a scanning line.

Datasets and Implementation

The framework was evaluated using massive 2D grayscale EM stacks with a resolution of 3–5 nm per pixel. Expert-annotated tiles served as the ground truth for evaluating the detection of cell membranes and organelles. Because the features are based on relative structural units (knots) rather than absolute pixel values, they demonstrate significant robustness to variations in chemical staining.

Figure 3: Representative EM image tiles (a, b, c) showing the high-resolution neural tissue and the specific structures (colored boxes) targeted by the RLF framework.

Key Results

RLFs significantly outperformed traditional Gabor filters and Hessian-based diffusion methods in several key areas:

  • Cell Boundary Enhancement: RLFs produced cleaner, crisper boundaries compared to traditional diffusion methods.
  • Mitochondria Segmentation: ROC analysis showed that RLFs achieved higher true-positive rates with fewer false alarms than Texton-based classifiers.
  • Computational Efficiency: The algorithm is "embarrassingly parallel." On a 64-core cluster, a 512 x 512 image was processed across 180 directions in just 11 seconds.

Conclusion and Future Work

RLFs provide a versatile and scalable tool for the automated analysis of neural tissue. While currently dependent on initial edge detection for knot placement, future work may involve utilizing the full multi-dimensional feature distribution and integrating more sophisticated classifiers like Support Vector Machines (SVM) to further enhance segmentation accuracy across diverse connectomic datasets.

Summary updated

BibTeX
@inproceedings{kumar2010radon,
  author = {Ritwik Kumar and Amelio Vázquez Reina and Hanspeter Pfister},
  title = {{Radon-Like features and their application to connectomics}},
  booktitle = {Workshops @ CVPR},
  year = {2010},
  doi = {10.1109/cvprw.2010.5543594},
}
Figure