Boosting Hand-Crafted Features for Curvilinear Structure Segmentation by Learning Context Filters

Emanuele Trucco, Frsa, Fiapr, Fbmva, Faaia, Roberto Annunziata, Ahmad Kheirkhah, Pedram Hamrah, Emanuele TruccoMICCAI 2015 · DOI

TL;DR

This paper introduces a novel method for segmenting curvilinear structures (like blood vessels) by combining traditional hand-crafted features with learned "context filters." This approach significantly improves segmentation accuracy and computational efficiency compared to existing methods, particularly for challenging datasets with low signal or complex structures.

What Problem Does This Paper Address?

Accurate segmentation of curvilinear structures in medical images is crucial for diagnosing various diseases. These structures, such as blood vessels, nerve fibers, and neurons, present significant segmentation challenges due to low signal-to-noise ratios, non-uniform illumination, and complex configurations like bifurcations. While hand-crafted features (HCFs) are fast, they can be brittle. Fully learned methods are powerful but computationally expensive. Recent hybrid approaches combine HCFs with learned appearance filters, but this paper argues that these learned filters can be redundant and don't fully capture the contextual relationships between structures, leading to false positives and missed segments, especially in fragmented or weakly connected cases. The authors aim to improve segmentation by incorporating context information into the learned filters, without sacrificing computational efficiency.

Methods

The core of the proposed method is a novel combination of existing hand-crafted features (specifically, the Optimally Oriented Flux or OOF feature) with newly learned context filters. Unlike previous hybrid methods that learn appearance filters, these new filters are trained to capture relationships between structures.

The learning process for these context filters uses unsupervised k-means clustering. Patches from the input images are pre-processed (whitened) and then mapped to a dictionary of learned filters. This process aims to find a sparse representation of the data. The key innovation is that these context filters are learned after the initial appearance features (like OOF) have been extracted, allowing them to leverage the information already encoded in the HCFs.

The final description vector for each pixel combines the OOF feature (representing appearance) with the outputs of the learned context filters. This combined vector is then fed into a Random Decision Forest classifier for pixel-level segmentation.

Datasets

The method was validated on three challenging datasets:

  • IVCM: Images of corneal nerve fibers, characterized by low resolution, non-uniform contrast, tortuosity, and fragmentation.
  • BF2D: High-resolution micrographs of neurons, featuring low signal-to-noise and point-like structures that can be mistaken for noise.
  • VC6: Images of dendritic and axonal subtrees from neurons, known for artifacts and poor contrast.

The authors used standard dataset partitions where available and performed random sub-sampling for cross-validation.

Key Insights and Ideas

  • Context is King (for Segmentation): The paper strongly suggests that incorporating contextual information (how structures relate to each other) is more crucial for robust curvilinear segmentation than just learning more appearance filters.
  • Learned Filters Can Be Redundant: The authors highlight that learned appearance filters can often be reconstructed from existing HCFs, diminishing their unique contribution. Their method implicitly reduces this redundancy.
  • K-means is Sufficient and Fast: For learning filters, k-means clustering proved to be a fast and effective alternative to more complex methods like sparse coding, especially when dealing with large dictionaries.
  • Efficiency Gains: By learning context filters on a different "layer" (after HCFs) and using k-means, the method achieves competitive or superior results with significantly reduced computational cost compared to prior work.

Key Results

  • The proposed method ("OOF, learned context") consistently outperformed baseline methods (including a recent hybrid approach using learned appearance filters) across all three datasets.
  • Performance gains were particularly notable on the IVCM and VC6 datasets, which feature more fragmented and complex structures.
  • The method showed improved precision at low recall for fragmented structures and higher precision at high recall for point-like structures, effectively reducing false positives.
  • Learning a smaller number of context filters (e.g., 10) often yielded comparable or better results than learning a larger number of appearance filters (e.g., 100), demonstrating reduced redundancy.
  • The k-means approach for learning context filters was significantly faster than sparse coding, with training times measured in minutes rather than days for large filter banks.

Limitations and Future Work

  • Limitations: While the paper focuses on the benefits of learned context filters, it acknowledges that learned appearance filters using sparse coding can offer slightly better performance (though at a much higher computational cost). The current work primarily uses k-means for learning.
  • Future Work: The authors plan to investigate combining learned context filters with their recently proposed HCFs and explore designing new HCFs that can optimally exploit information from the appearance model.

Summary updated

BibTeX
@inproceedings{faaia2015boosting,
  author = {Emanuele Trucco, FRSA, FIAPR, FBMVA, FAAIA and Roberto Annunziata and Ahmad Kheirkhah and Pedram Hamrah and Emanuele Trucco},
  title = {{Boosting Hand-Crafted Features for Curvilinear Structure Segmentation by Learning Context Filters}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2015},
  doi = {10.1007/978-3-319-24574-4\_71},
}
Figure