Multi-scale Nodule Detection in Chest Radiographs

Arnold M. R. Schilham, Bram van Ginneken, Marco LoogMICCAI 2003 · OpenReview

Abstract

Early detection is the most promising way to enhance a patient’s chance for survival of lung cancer. In this work, a novel computer algorithm for nodule detection in chest radiographs is presented that takes into account the wide size range for lung nodules through the use of multi-scale image processing techniques. The method consists of: i) Lung field segmentation with an Active Shape Model; ii) Nodule candidate detection by Lindeberg’s multi-scale blob detector and quadratic classification; iii) Blob segmentation by multi-scale edge focusing; iv) k-Nearest neighbor classification. Experiments on the complete JSRT database show that by accepting on average 2 false positives per image, 50.6% of all nodules are detected. For 10 false positives, this increases to 69.5%.

Introduction

Lung cancer remains a leading cause of mortality worldwide, making early detection of pulmonary nodules essential for improving patient outcomes. However, identifying these nodules on chest X-rays is a significant challenge for radiologists. Nodules are often subtle, vary greatly in size (ranging from 6mm to 60mm), and can be obscured by overlapping anatomical structures such as ribs, the heart, or the diaphragm. This paper introduces an automated computer-aided diagnosis (CAD) system designed to assist radiologists by flagging potential nodule locations, specifically addressing the multi-scale nature of the detection problem.

Methods

The proposed system utilizes a multi-stage pipeline to process radiographs and identify candidates.

Flowchart of the proposed multi-stage nodule detection pipeline.

Pipeline Stages

  1. Preprocessing and Segmentation: Images are subsampled to 1024x1024 pixels. An Active Shape Model (ASM) is employed to segment the lung fields, restricting the search area to reduce false positives from non-pulmonary regions.
  2. Local Normalization (LN): To handle the high dynamic range of X-rays, a local normalization filter equalizes contrast, revealing nodules that might be hidden in very bright or dark areas.
  3. Multi-scale Blob Detection: The system uses Lindeberg’s multi-scale blob detector to find bright, circular structures across various scales ($\sigma = 1$ to $16$ pixels).
  4. Initial Candidate Selection: A quadratic classifier using 13 features (e.g., Hessian matrix components, intensity) filters the initial detections down to approximately 20 candidates per image.
  5. Refined Segmentation: Scale-space edge focusing is used to accurately define nodule boundaries. By tracing edges from coarse to fine scales along radial rays, the system better separates nodules from the background.
  6. Final Classification: A k-Nearest Neighbor (kNN) classifier evaluates 98 features (eventually pruned to 13) to determine the final probability of a candidate being a nodule.
Visual progression of the algorithm: (a) original image, (b) local normalization, (c) initial blob detection, and (d) candidate selection.

Results and Discussion

The algorithm was evaluated using the JSRT database, which includes 154 radiographs with proven nodules of varying subtlety.

  • Detection Performance: The system achieved a sensitivity of 50.6% at 2 false positives (FP) per image, increasing to 69.5% at 10 FP.
  • Subtlety Analysis: While performance was highest on "obvious" nodules, the system successfully flagged nearly 30% of "extremely subtle" cases at 4 FP per image—cases that are frequently missed by human observers.
  • Limitations: Approximately 8.5% of nodules were missed because they were located behind the heart or diaphragm, regions currently excluded by the ASM segmentation.

Conclusion

By integrating multi-scale blob detection with contrast normalization and refined segmentation, this CAD system effectively addresses the size variability of lung nodules. Future work will focus on expanding the segmentation to include retrocardiac and subdiaphragmatic areas and refining feature engineering to further reduce false positives in a clinical setting.

Summary updated

Abstract

Early detection is the most promising way to enhance a patient’s chance for survival of lung cancer. In this work, a novel computer algorithm for nodule detection in chest radiographs is presented that takes into account the wide size range for lung nodules through the use of multi-scale image processing techniques. The method consists of: i) Lung field segmentation with an Active Shape Model [1]; ii) Nodule candidate detection by Lindeberg’s multi-scale blob detector [2] and quadratic classification; iii) Blob segmentation by multi-scale edge focusing; iv)k Nearest neighbor classification. Experiments on the complete JSRT database [3] show that by accepting on average 2 false positives per image, 50.6% of all nodules are detected. For 10 false positives, this increases to 69.5%.
BibTeX
@inproceedings{schilham2003multi,
  author = {Arnold M. R. Schilham and Bram van Ginneken and Marco Loog},
  title = {{Multi-scale Nodule Detection in Chest Radiographs}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2003},
  url = {https://openreview.net/forum?id=CnSgSdpNne},
}
Figure