Automatic Detection Of Rib Contours in Chest Radiographs

TL;DR

This paper presents an automated algorithm for detecting rib contours in chest radiographs, a crucial step for improving the accuracy of tumor detection and providing anatomical reference. The method employs a multi-stage approach involving image preprocessing, local edge detection, global boundary fitting with conic sections and line segments, and a heuristic refinement process.

What Problem Does This Paper Address?

Detecting rib contours in chest radiographs automatically is a challenging task due to several factors:

  • Variable Optical Density: The brightness of rib images changes significantly along their length, and parts of contours can be obscured.
  • Intersections and Superimpositions: Ribs often intersect with each other and with other anatomical structures like blood vessels and tumors, making individual contour identification difficult.
  • False Positives in Tumor Detection: Rib images can be mistakenly identified as tumors by automated detection systems, leading to false positives.
  • Lack of Anatomical Reference: Precise rib contours are needed to establish a reliable frame of reference for describing the location and extent of lesions within the thoracic cavity and for accurately computing the lung boundary.

Existing methods were found to be limited, often restricted to specific regions of the radiograph, degraded by overlapping tumors, or failing to detect lower ventral rib contours. This paper aims to overcome these limitations with a more robust and comprehensive algorithm.

Methods

The proposed system is a four-step pipeline:

  1. Preprocessing (Step 1): The input chest radiograph is digitized and then subjected to a high-pass digital filter. This step enhances high spatial frequencies, which are primarily associated with edges and sharp changes in gray-level density, making the rib structures more distinct. The filtering is performed in the spatial frequency domain using a specific transfer function.
  2. Local Edge Detection (Step 2): This stage identifies potential edge elements. It combines three operators:
  3. Gradient: Computes the modulus and direction of the spatial gradient to indicate the presence and orientation of edges.
  4. Laplacian: Calculates the Laplacian to detect edges and is sensitive to local gray-level changes.
  5. Thresholding: Thresholds are applied to both the gradient modulus and the Laplacian to select reliable edge elements, filtering out noise. A binary edge map is generated.
  6. Global Boundary Detector (Step 3): This step fits global models to the detected edge elements to form continuous contours.
  7. Ventral Ribs: Straight line segments are used to approximate the contours of the ventral ribs. A Hough transform-like accumulator array procedure is employed to detect these line segments.
  8. Dorsal Ribs: Conic sections (parabolas and ellipses) are used to approximate the contours of the dorsal ribs. The algorithm extends the Hough technique to detect conic sections by identifying potential foci in a parameter space.
  9. Refinement of Rib Boundaries (Step 4): This final stage refines the contours obtained in Step 3.
  10. Strips and Heuristic Search: Small "strips" are created centered around the line segments and conic sections identified in Step 3. These strips act as a "plan" for finding the actual rib contours.
  11. Heuristic Function: A heuristic function h(i,j) is defined, incorporating gradient and Laplacian information, to guide a search within these strips. This function penalizes deviations from the expected contour.
  12. Smoothing: A smoothing algorithm is applied to the refined contours to produce the final, smooth rib boundaries. This step utilizes a look-ahead mechanism in a tree-like search, inspired by hierarchical abstraction spaces.

Datasets

The paper describes the processing of digitized chest radiographs. The input images are initially reduced photographically and then digitized into a 1024x896 array of 8-bit pixels. A consolidation operation reduces this to 256x224 pixels. The paper does not explicitly name or refer to a specific public dataset. The figures show examples of processed radiographs, implying the use of a collection of chest X-ray images.

Key Insights and Ideas

  • Multi-stage Approach: The success of the algorithm relies on a carefully designed sequence of steps, where each stage addresses specific challenges in rib contour detection.
  • Combining Local and Global Information: The algorithm effectively combines local edge detection with global boundary fitting (lines and conic sections) to achieve robust contour extraction.
  • Hough Transform Extension: Extending the Hough transform concept to detect conic sections for dorsal ribs is a novel approach for this specific application.
  • Heuristic Search with Look-Ahead: The refinement step's use of a heuristic function and a look-ahead mechanism within defined strips is a sophisticated way to handle noise and gaps in edge detection, drawing inspiration from AI planning techniques.
  • Hierarchical Abstraction: The algorithm embodies the concept of hierarchical abstraction, starting with a rough plan (strips) and progressively introducing details and refining the contours.

Key Results

The paper demonstrates the effectiveness of the algorithm through qualitative results shown in figures.

  • The algorithm successfully detects rib contours, including those partially obscured by other structures like tumors.
  • The final results (Figures 8 and 9) show smooth, well-defined dorsal and ventral rib contours that closely match human perception.
  • The system is designed to be implementable on a small computer with limited memory, suggesting computational efficiency.
  • The paper claims its algorithm overcomes weaknesses of prior work, such as degradation by tumors and failure to detect lower ventral rib contours.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that some portions of the detected contours might be "overly thick roads" or have gaps before the refinement step.
  • While the algorithm can handle some obscuration by tumors, the extent of this capability might have limits.
  • The paper mentions that the ventral rib detection stops when costal cartilages are expected, implying a boundary to its detection.
  • Future Work:
    • Texture Analysis: Incorporating texture analysis of lung tissue and bone structure could further improve detection.
  • World-Model Representation: Building an explicit world-model representation of the rib cage could enhance understanding and detection.
  • Computer-Radiologist Interaction: Developing interactive tools for computer-radiologist collaboration could refine the system and its clinical utility.

Summary updated

BibTeX
@inproceedings{1975automatic,
  title = {{Automatic Detection Of Rib Contours in Chest Radiographs}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1975},
}
Figure