Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Foreign object segmentation in chest x-rays through anatomy-guided shape insertion

Constantin Seibold, Hamza Kalisch, Lukas Heine, Simon Reiß, Jens Kleesiek

TL;DR

This work addresses the problem of segmenting foreign objects in chest X-rays, hindered by high object diversity and limited labeled data. It introduces an anatomy-guided synthetic data pipeline that combines Structure Plotting and Cut-Paste augmentation anchored to anatomical segmentation to generate ground-truth FB masks. Evaluations across state-of-the-art instance segmentation models show that Mask2Former trained on synthetic data can match fully supervised performance while using up to 93% fewer manual annotations, and can transfer effectively to real FB datasets and unseen X-ray domains. The approach offers a scalable pathway for semi-automatic large-scale dataset annotation and improved FB segmentation in clinical radiography.

Abstract

In this paper, we tackle the challenge of instance segmentation for foreign objects in chest radiographs, commonly seen in postoperative follow-ups with stents, pacemakers, or ingested objects in children. The diversity of foreign objects complicates dense annotation, as shown in insufficient existing datasets. To address this, we propose the simple generation of synthetic data through (1) insertion of arbitrary shapes (lines, polygons, ellipses) with varying contrasts and opacities, and (2) cut-paste augmentations from a small set of semi-automatically extracted labels. These insertions are guided by anatomy labels to ensure realistic placements, such as stents appearing only in relevant vessels. Our approach enables networks to segment complex structures with minimal manually labeled data. Notably, it achieves performance comparable to fully supervised models while using 93\% fewer manual annotations.

Foreign object segmentation in chest x-rays through anatomy-guided shape insertion

TL;DR

This work addresses the problem of segmenting foreign objects in chest X-rays, hindered by high object diversity and limited labeled data. It introduces an anatomy-guided synthetic data pipeline that combines Structure Plotting and Cut-Paste augmentation anchored to anatomical segmentation to generate ground-truth FB masks. Evaluations across state-of-the-art instance segmentation models show that Mask2Former trained on synthetic data can match fully supervised performance while using up to 93% fewer manual annotations, and can transfer effectively to real FB datasets and unseen X-ray domains. The approach offers a scalable pathway for semi-automatic large-scale dataset annotation and improved FB segmentation in clinical radiography.

Abstract

In this paper, we tackle the challenge of instance segmentation for foreign objects in chest radiographs, commonly seen in postoperative follow-ups with stents, pacemakers, or ingested objects in children. The diversity of foreign objects complicates dense annotation, as shown in insufficient existing datasets. To address this, we propose the simple generation of synthetic data through (1) insertion of arbitrary shapes (lines, polygons, ellipses) with varying contrasts and opacities, and (2) cut-paste augmentations from a small set of semi-automatically extracted labels. These insertions are guided by anatomy labels to ensure realistic placements, such as stents appearing only in relevant vessels. Our approach enables networks to segment complex structures with minimal manually labeled data. Notably, it achieves performance comparable to fully supervised models while using 93\% fewer manual annotations.
Paper Structure (12 sections, 4 figures, 1 table)

This paper contains 12 sections, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Anatomy-Guided Structure Insertion
  3. Structure Plotting
  4. Cut-Paste of Pre-extracted FB Crops
  5. Experimental Setup
  6. Implementation Details
  7. Datasets
  8. Evaluation
  9. Performance on Synthetic Data
  10. Transfer of Models
  11. Discussion and Conclusion
  12. Compliance with ethical standards

Figures (4)

  • Figure 1: Left: We display some plotted structures using anatomical guidance. From top left to bottom right as follows: Parallel lines from the clavicle to a ventricle, surgical clips at a rib, a grid structure at the trachea, a general line. Right: We display a subset for the different more composite foreign objects.
  • Figure 2: Pseudo Foreign Object insertion via Anatomical Guidance. After segmenting anatomical structures of a given CXR, we sample either a cutout or a synthetic structure to be plotted. Using the sample, we then simulate a foreign body at the corresponding anatomical position by either inserting or plotting the structure.
  • Figure 3: We show the mAP performances of the five considered baselines across different artificial dataset sizes on the synthetic validation set of SynthFB.
  • Figure 4: We can see that Mask2Former trained on SynthFB is able to also segment elements in other X-ray domainsncbi_image.