TL;DR
This paper introduces SCIRD, a novel ridge detector that is invariant to scale, rotation, and curvature. It overcomes limitations of existing methods by relaxing assumptions about local structure linearity, leading to improved segmentation of tortuous and fragmented curvilinear structures in challenging imaging scenarios.
What Problem Does This Paper Address?
Accurately segmenting curvilinear structures like blood vessels, dendritic trees, and nerve fibers is crucial for various medical imaging applications, such as lesion quantification and therapy monitoring. Many existing methods rely on hand-crafted features (HCFs) that assume local linearity and tubularity. However, these assumptions are often violated by real-world structures that are tortuous, fragmented, or exhibit varying widths, especially in low-resolution or noisy images. This paper aims to develop a more robust ridge detector that can handle these complexities.
Methods
The core of the SCIRD method is a novel "curved-support Gaussian model." This model extends the traditional Gaussian by incorporating a non-linear transformation that allows its support to bend, effectively mimicking curved structures. The tubularity measure is then computed by convolving the image with a filter bank derived from second-order directional derivatives of these curved-support Gaussians.
Key innovations include:
- Curved-Support Gaussian Model: This allows the detector to adapt to non-linear structures, unlike traditional linear models.
- Scale and Curvature Invariance: The detector is designed to be invariant to these properties, making it more robust to variations in image acquisition and structure appearance.
- Filter Bank Generation: To achieve rotation invariance and handle fragmentation, SCIRD uses a bank of kernels with varying orientations and parameters that control the "memory" (elongation) of the Gaussian support.
- Contrast Normalization: An optional contrast normalization term is introduced to enhance or suppress structures based on local image contrast, further improving performance in challenging conditions.
- Supervised Extension: SCIRD can be integrated with a supervised classifier (Random Decision Forest) to further improve segmentation accuracy.
Datasets
The SCIRD method was evaluated on three challenging datasets:
- IVCM: Confocal microscopy images of the corneal subbasal nerve plexus. This dataset is characterized by low resolution, non-uniform contrast, tortuosity, and fragmentation.
- BF2D: Minimum intensity projections of bright-field micrographs of neurons. These images have high resolution but low signal-to-noise ratio, with dendrites often appearing as fragmented, point-like structures.
- VC6: 3D images of dendritic and axonal subtrees from neurons, featuring artifacts and poor contrast.
Key Insights and Ideas
- Relaxing Linearity is Key: The paper strongly demonstrates that moving away from the assumption of locally straight structures is critical for accurately segmenting complex biological networks.
- Curvature Invariance Matters: Explicitly designing for curvature invariance allows the detector to perform well on highly tortuous paths where traditional methods fail.
- Contrast Normalization as a Powerful Tool: The ability to either enhance or suppress structures based on local contrast provides significant flexibility and performance gains, particularly for images with varying illumination or where certain structures are less relevant.
- Unsupervised Performance is Strong: Even without supervised learning, SCIRD shows competitive results, highlighting the effectiveness of its hand-crafted features.
Key Results
- SCIRD outperforms state-of-the-art hand-crafted feature (HCF) based ridge detectors (Frangi, Gabor, OOF) on all three challenging datasets.
- Qualitative results (Figure 3) show SCIRD provides better connectivity and higher signal-to-noise ratio in tubularity estimation.
- Quantitative results (Figure 4) show SCIRD achieves higher precision at medium to high recall values, especially on the IVCM dataset, indicating better handling of low resolution and low SNR with tortuous/fragmented structures.
- The contrast enhancement option ("SCIRD a ≠ 0") significantly boosts performance on the IVCM dataset, achieving results comparable to supervised methods.
- For BF2D, unsupervised SCIRD shows significant improvement, and contrast reduction ("SCIRD a ≠ 0") is beneficial.
- Supervised SCIRD further improves performance across all datasets.
- The method is computationally efficient, with run times on an Intel i7-4770 CPU being 0.86s (IVCM), 0.62s (VC6), and 8.75s (BF2D), and is highly parallelizable.
Limitations and Future Work
- Limitations: The paper notes that the initial formulation of the curved-support Gaussian model requires prior knowledge of structure orientation, though this is addressed by using multiple rotated kernels. The performance on the VC6 dataset did not benefit from contrast enhancement.
- Future Work: The authors plan to investigate combining SCIRD with learned filters to capture structures that are difficult to model, such as crossings and artifacts.
Summary updated