Automatic Detection Of Rib Contours in Chest Radiographs

Harry Wechsler, Jack Sklansky

Introduction

Automated analysis of chest radiographs has historically been complicated by the presence of rib structures. While ribs serve as vital anatomical landmarks, their overlap with soft tissue, blood vessels, and potential pathologies creates significant challenges for computer vision. Specifically, rib shadows can lead to false-positive tumor detections, and the lack of a precise rib "map" hinders the accurate localization of lesions or the definition of lung boundaries.

In their 1974 work, Wechsler and Sklansky proposed a foundational multi-stage pipeline to automatically segment these structures. Detecting rib contours is inherently difficult due to varying bone density along the curve and frequent obscuration by other anatomical features or pathologies.

Methodology

The proposed system utilizes a four-step algorithmic approach designed to operate within the severe memory constraints of 1970s hardware.

Block diagram of the rib contour detector pipeline.
  1. Preprocessing: Digitized images are subjected to high-pass filtering in the spatial frequency domain via Fast Fourier Transform (FFT). This step sharpens edges and normalizes density variations across the lung field.
  2. Local Edge Detection: A combination of Gradient and Laplacian operators is applied. Edge elements are filtered based on magnitude and angular orientation, resulting in a "direction map" of potential rib boundaries.
  3. Global Boundary Detection: This stage employs a modified Hough Transform. While typically used for lines, the authors adapted it to detect conic sections (parabolas and ellipses) for dorsal ribs and straight line segments for ventral ribs. An accumulator array identifies the most probable geometric parameters.
  4. Boundary Refinement: A heuristic search with look-ahead evaluates potential paths within a narrow "plan" strip defined by the global detector. The algorithm selects the path that maximizes a specific heuristic function, effectively refining the coarse geometric fit.

Data and Implementation

The study utilized standard chest radiographs (100 mm x 125 mm) digitized into a 1024 x 896 array of 8-bit pixels. Due to the limited 8,000-word core memory of the minicomputer used, images were downsampled to a final resolution of 256 x 224 pixels for processing.

Results and Insights

The algorithm demonstrated significant robustness, particularly in its ability to "bridge the gap" when ribs were partially obscured by tumors—a feat made possible by the use of global geometric models.

Key findings included:

  • Anatomical Differentiation: The system successfully accounted for the differences between dorsal ribs (concave downward, more distinct) and ventral ribs (concave upward, fainter).
  • Detection Accuracy: In testing, the Hough detector identified 21 out of 26 ventral rib contours directly. The remaining five were resolved using thickness-based extrapolation and spatial context from neighboring ribs.
  • Qualitative Performance: Refined contours showed a high degree of correlation with human perception, even in complex areas where dorsal and ventral rib images intersected.

Conclusion

Despite its success, the authors noted limitations in the high-structural density of the hilar region. They suggested that future iterations could benefit from 3D spatial world-models and texture analysis to better distinguish bone from surrounding tissue. This work remains a landmark in medical imaging, demonstrating sophisticated global-to-local search strategies under extreme computational constraints.

Summary updated

BibTeX
@inproceedings{wechsler1975automatic,
  author = {Harry Wechsler and Jack Sklansky},
  title = {{Automatic Detection Of Rib Contours in Chest Radiographs}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1975},
}
Figure