Multiscale detection of filamentary features in image data

Xiaoming Huo, Jihong Chen, David L. DonohoProceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE 2003 · DOI

TL;DR

This paper introduces a novel multiscale statistical framework for detecting faint, curvilinear "filamentary" features in noisy digital images. By adapting concepts from point process analysis to image data and introducing "digital axoids," the method offers a theoretically optimal and practically robust way to identify these structures, which are common in fields like medical imaging and satellite analysis.

What Problem Does This Paper Address?

Detecting subtle, elongated structures like filaments in noisy images is a fundamental challenge across various scientific and engineering disciplines, including medical image analysis, target recognition, and satellite imagery. While previous statistical methods have shown promise for detecting straight line segments, they fall short when dealing with the more common, and often fainter, curvilinear features. The core problem is to develop a robust and statistically sound method that can reliably identify these filamentary structures even when they are weak and embedded in significant noise, and to do so efficiently in a digital image context.

Methods

The paper proposes a multiscale approach inspired by statistical methods for point processes. The key innovation is the introduction of "digital axoids," which are essentially thickened digital lines. These axoids act as basic building blocks that can cover features at various scales, orientations, and positions.

The detection process involves:

  1. Digital Axoids: These are defined as sets of pixels that form a thickened digital line, generated using Bresenham's line-drawing algorithm and then expanded to a certain thickness. They are parameterized by scale ($s$), strip index ($k$), starting row ($l_0$), and slope parameter ($l_1$).
  2. Significance Graph: For each digital axoid, a statistic ($X(s, k, l_0, l_1)$) is computed. This statistic measures the average intensity within the axoid. If this statistic exceeds a threshold ($N^*$), the axoid is considered "significant." A graph is then constructed where significant axoids are vertices, and edges connect axoids that are geometrically connected and have similar slopes ("good continuation").
  3. Longest Run: The core detection statistic is the length of the "longest run" ($|R(s)|$) in the significance graph at each scale. A longest run is a chain of connected significant axoids.
  4. Hypothesis Testing: The method frames the problem as a hypothesis test: $H_0$ (no feature, only noise) vs. $H_a$ (an underlying feature exists). The overall length of the longest run across all scales ($L^0 = \max_s |R(s)|$) is compared against a threshold ($L^*$) to make a decision. If $L^0 > L^*$, the null hypothesis is rejected, indicating the presence of a feature.

The paper also details how the parameters $N^*$ and $L^*$ are calibrated, drawing on theoretical optimality and simulation results.

Datasets

The paper uses synthetic data generated to represent various types of filamentary features. These include:

  • Trigonometric Functions: Features defined by sinusoidal curves.
  • Thickened Trigonometric Functions: Features created by adding adjacent pixels to the trigonometric curves.
  • Beams: Features based on linear functions.
  • Thickened Beams: Features created by adding adjacent pixels to the linear beams.

These synthetic datasets are used to evaluate the performance of the proposed detection method under varying noise levels and feature amplitudes. No real-world datasets are used for evaluation in this paper.

Key Insights and Ideas

  • Bridging Continuous and Digital: A significant conceptual leap is the adaptation of continuous geometric concepts (like Lipschitz conditions) and statistical frameworks (from point processes) to the discrete, pixelated world of digital images.
  • Digital Axoids as Flexible Building Blocks: The introduction of "digital axoids" provides a flexible way to represent and detect features of varying scales, orientations, and thicknesses, overcoming limitations of methods designed for straight lines.
  • Longest Run as a Powerful Detector: The length of the longest chain of "significant" axoids (the longest run) proves to be a highly sensitive indicator of underlying filamentary structures, especially when compared to random noise.
  • Multiscale Analysis is Crucial: The effectiveness hinges on analyzing features across multiple scales, as a feature might be best captured by axoids of a specific size and orientation.
  • Theoretical Optimality Meets Practicality: The method is grounded in statistical theory, aiming for asymptotically optimal detection, while also being designed for practical implementation and demonstrating effectiveness in simulations.

Key Results

  • The proposed multiscale detection method is shown to be effective in identifying underlying filamentary features in synthetic noisy images.
  • Simulations demonstrate that the detection rate increases significantly with the amplitude ($A_n$) of the underlying feature. For instance, with $n=64$ and $t=1$, a detection rate of 70% is achieved with $A_n = 2.666$, and 100% with $A_n = 3.000$.
  • The method is capable of detecting features even when they are "obscure" to the human eye, as illustrated in Figure 8 where a feature with $A_n = 0.5$ is detected with nearly 100% accuracy on a 128x128 image.
  • The length of the longest run is substantially larger when an underlying feature is present compared to pure noise, as visualized in Figure 10.
  • The computational cost increases with image size ($n$), with a notable jump from $n=128$ to $n=256$ (Table 3).

Limitations and Future Work

  • Thickness Limitation: The paper acknowledges that the presented axoid definition primarily handles features with a height of 1 and suggests generalizing this to thicker features as an exercise for the reader.
  • Computational Cost: While the method is theoretically efficient, the computational time increases significantly with image resolution, as shown in Table 3.
  • Parameter Calibration: The calibration of thresholds ($N^*$ and $L^*$) relies on theoretical considerations and simulations, and fine-tuning these parameters might be necessary for specific applications.
  • Future Work: The authors suggest exploring variations in the design of digital axoids and their impact on detection efficiency and sensitivity. Further investigation into the theoretical optimality and practical performance across a wider range of feature types and noise conditions would also be valuable.

Summary updated

Abstract

Taking advantage of the new developments in mathematical statistics, a multiscale approach is designed to detect filament or filament-like features in noisy images. The major contribution is to introduce a general framework in cases when the data is digital. Our detection method can detect the presence of an underlying curvilinear feature with the lowest possible strength that are still detectible in theory. Simulation results on synthetic data will be reported to illustrate its effectiveness in finite digital situations.
BibTeX
@inproceedings{huo2003multiscale,
  author = {Xiaoming Huo and Jihong Chen and David L. Donoho},
  title = {{Multiscale detection of filamentary features in image data}},
  booktitle = {Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE},
  year = {2003},
  doi = {10.1117/12.503721},
}

Loading related content...

Figure