Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Hierarchical Loss And Geometric Mask Refinement For Multilabel Ribs Segmentation

Aleksei Leonov, Aleksei Zakharov, Sergey Koshelev, Maxim Pisov, Anvar Kurmukov, Mikhail Belyaev

TL;DR

This work tackles automatic multilabel rib segmentation and numeration in CT scans by introducing a hierarchical loss that jointly optimizes rib binary segmentation and rib-type classification within a single end-to-end model. A two-headed U-Net architecture is augmented with a binary rib-mask head and a 12-class rib-type head, guided by a loss term that selectively penalizes classification only in rib voxels, and complemented by a geometric mask refinement postprocessing step to improve lower-rib labeling. Using the public RibSeg v2 dataset, the approach achieves state-of-the-art performance with a peak average label-accuracy of $98.2\%$, significantly outperforming prior methods and demonstrating robustness on challenging cases. The work also provides a critical audit of RibSeg v2 annotations to stress-test evaluation fairness and highlights avenues for annotation improvements to further boost model quality. Overall, the proposed method offers a fast, end-to-end solution for rib segmentation and labeling with practical implications for radiology workflows.

Abstract

Automatic ribs segmentation and numeration can increase computed tomography assessment speed and reduce radiologists mistakes. We introduce a model for multilabel ribs segmentation with hierarchical loss function, which enable to improve multilabel segmentation quality. Also we propose postprocessing technique to further increase labeling quality. Our model achieved new state-of-the-art 98.2% label accuracy on public RibSeg v2 dataset, surpassing previous result by 6.7%.

Hierarchical Loss And Geometric Mask Refinement For Multilabel Ribs Segmentation

TL;DR

This work tackles automatic multilabel rib segmentation and numeration in CT scans by introducing a hierarchical loss that jointly optimizes rib binary segmentation and rib-type classification within a single end-to-end model. A two-headed U-Net architecture is augmented with a binary rib-mask head and a 12-class rib-type head, guided by a loss term that selectively penalizes classification only in rib voxels, and complemented by a geometric mask refinement postprocessing step to improve lower-rib labeling. Using the public RibSeg v2 dataset, the approach achieves state-of-the-art performance with a peak average label-accuracy of , significantly outperforming prior methods and demonstrating robustness on challenging cases. The work also provides a critical audit of RibSeg v2 annotations to stress-test evaluation fairness and highlights avenues for annotation improvements to further boost model quality. Overall, the proposed method offers a fast, end-to-end solution for rib segmentation and labeling with practical implications for radiology workflows.

Abstract

Automatic ribs segmentation and numeration can increase computed tomography assessment speed and reduce radiologists mistakes. We introduce a model for multilabel ribs segmentation with hierarchical loss function, which enable to improve multilabel segmentation quality. Also we propose postprocessing technique to further increase labeling quality. Our model achieved new state-of-the-art 98.2% label accuracy on public RibSeg v2 dataset, surpassing previous result by 6.7%.
Paper Structure (17 sections, 1 equation, 2 figures, 3 tables)

This paper contains 17 sections, 1 equation, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Architecture
  5. Hierarchical loss function
  6. Geometric mask refinement
  7. Data and Metrics
  8. RibSeg v2 Dataset
  9. Metrics
  10. Experimental setup
  11. Pipeline
  12. Training
  13. Inference
  14. Results
  15. Conclusion
  16. ...and 2 more sections

Figures (2)

  • Figure 1: Postprocessing algorithm. For simplicity, only right ribs are shown. Left: initial prediction. Middle: connected components are shown with alternating red and blue. Numbers indicate "probable" ribs types. Right: final result. Untouched components are transparent.
  • Figure 2: Filtration of the ribs near the spine. Center and right columns: coronal plain - annotation and prediction masks with filtering respectively. Red color indicates ribs mask, green color shows area where we filter masks to mitigate the impact of inaccurate annotation, blue color denotes filtered part of ribs.