Automatic Localization of the Lumbar Vertebral Landmarks in CT Images with Context Features

Dimitrios Damopoulos, Ben Glocker, Guoyan ZhengMSKI @ MICCAI 2017 · DOI

TL;DR

This paper introduces a novel two-layer approach for automatically pinpointing anatomical landmarks on lumbar vertebrae in CT scans. By incorporating "context features" from neighboring vertebrae, the method significantly improves localization accuracy, addressing a key challenge in automated spinal analysis and diagnosis.

What Problem Does This Paper Address?

Low back pain is a widespread health issue, and accurately analyzing the lumbar spine is crucial for diagnosis and treatment. This often involves identifying specific anatomical landmarks on the vertebrae. However, the repetitive structure of the spine makes it difficult for automated systems to distinguish between adjacent vertebrae, leading to localization errors. Existing methods often struggle with this ambiguity, especially when relying solely on local image features. This paper aims to overcome this by developing a more robust landmark localization system that leverages information from the surrounding vertebral context.

Methods

The proposed framework consists of two main modules:

  1. Vertebral Body (VB) Localization and Pose Estimation: This module first performs a "first-level detection" of VB centers using a random forest classifier. Then, it employs a "Curved Planar Reformation" technique to reslice the image along a spline passing through these detected centers. A mean-intensity profile is generated from this resliced image, and local maxima (peaks) are identified as potential endplate locations. A k-Nearest Neighbors (k-NN) classifier is trained to distinguish true endplate peaks from false positives, allowing for accurate VB center estimation. The pose of each vertebra is derived from a Local Coordinate System (LCS) established during the spline reformation.
  2. Vertebral Landmark Localization (Two-Layered Approach): This module builds upon the VB centers and poses identified by the first module.
  3. First Layer (Appearance Features): For each lumbar level, a separate random forest regressor is trained to predict the displacement of specific landmarks from the VB center. This layer uses traditional Haar-like features based on local image intensity within a Region of Interest (ROI) around the VB center.
  4. Second Layer (Context Features): This layer enhances the first layer by incorporating "context features." These features are derived from distance maps generated by the first layer's predictions. To avoid bias, the method generates multiple displaced ROIs around the ground truth VB centers during training. The second layer's random forest then uses both the original appearance features and these new context features (derived from the distance maps) to refine the landmark predictions. This allows the model to learn from the spatial relationships and appearance of neighboring vertebrae.

Datasets

The methods were evaluated on a dataset of 28 lumbar-focused CT images. The paper does not mention introducing a new benchmark dataset.

Key Insights and Ideas

  • Context is King: The core insight is that local appearance alone is insufficient for robust landmark localization in the repetitive lumbar spine. Incorporating contextual information from neighboring vertebrae significantly improves accuracy.
  • Two-Layered Auto-context: The paper effectively adapts the "Auto-context" framework, typically used for segmentation, to a multi-landmark localization task. The two-layer approach allows for an iterative refinement of predictions.
  • Endplate-Based VB Center Detection: The proposed endplate detection method, using k-NN on mean-intensity profile peaks, provides a robust way to find VB centers, which are crucial for subsequent landmark localization.
  • Addressing Ambiguity: The second layer's ability to learn from distance maps helps disambiguate similar-looking structures across different vertebral levels, a common pitfall in single-layer approaches.

Key Results

  • VB Center Localization: The endplate-based VB localization method achieved a mean localization error of 3.2 mm with a standard deviation of 2.0 mm on the 8 testing images. All lumbar VBs were detected successfully (within 10 mm of manual annotation).
  • Landmark Localization: The proposed two-layer landmark localization method achieved a mean localization error of 3.0 mm with a standard deviation of 1.6 mm. Notably, 95.4% of the detected landmarks had a localization error of below 6 mm.
  • Bounding Box Overlap: The Dice overlap coefficient for bounding boxes of detected landmarks against manual annotations was 88.8% on average across all spinal levels.

Limitations and Future Work

  • Limited Dataset Size: The evaluation was performed on a relatively small dataset of 28 CT images (20 for training, 8 for testing).
  • Focus on Specific Landmarks: The current work focuses on VB centers and inferior articular processes. The authors suggest applying the method to other vertebral landmarks, such as endplate landmarks for spondylolisthesis measurement.
  • Future Directions:
    • Larger Datasets: Evaluating the methods on larger and more diverse datasets.
  • Improved Localization Accuracy: Investigating ways to further fine-tune detections, potentially using multi-scale approaches.
  • Inter-level Context: Exploring context features that span across different vertebral levels, not just immediate neighbors.
  • Handling Implants/Pathologies: The current dataset excluded implants, and cases with mild scoliosis and fractures were present but not specifically addressed for robustness. Future work could focus on handling these complexities.

Summary updated

BibTeX
@inproceedings{damopoulos2017automatic,
  author = {Dimitrios Damopoulos and Ben Glocker and Guoyan Zheng},
  title = {{Automatic Localization of the Lumbar Vertebral Landmarks in CT Images with Context Features}},
  booktitle = {MSKI @ MICCAI},
  year = {2017},
  doi = {10.1007/978-3-319-74113-0\_6},
}
Figure