PaperlayerView exact revision on arXiv

Source-linked AI summary

A Computed Tomography Vertebral Segmentation Dataset with Anatomical Variations and Multi-Vendor Scanner Data

Hans Liebl, David Schinz, Anjany Sekuboyina, Luca Malagutti, Maximilian T. Löffler, Amirhossein Bayat, Malek El Husseini, Giles Tetteh, Katharina Grau, Eva Niederreiter, Thomas Baum, Benedikt Wiestler, Bjoern Menze, Rickmer Braren, Claus Zimmer, Jan S. Kirschke

arXiv:2103.06360v1eess.IVcs.CV

TL;DR

Automated spine segmentation depends on extensive datasets with corresponding metadata, while rare anatomical variants are underrepresented and can challenge deep-learning algorithms. VerSe 2020 expands the dataset and challenge with variant-enriched CT images and hybrid-generated masks; the challenge results support training reliable, fully automated segmentation algorithms.

  • Problem

    Spine algorithms require extensive datasets with corresponding metadata, while rare anatomical variants are underrepresented in unselected training data and deep-learning algorithms fail in such cases.

    Source: Background and Summary · p. 3

  • Method

    VerSe 2020 released an expanded dataset through the second challenge, enriching it for rare variants and providing voxel-level masks prepared with a human-machine hybrid algorithm.

    Source: Background and Summary · pp. 3–4, Technical Validation · p. 11

  • Results

    Successful VerSe challenges in 2019 and 2020 confirm that reliable, fully automated deep-learning spine segmentation algorithms can be trained.

    Source: Background and Summary · p. 4

  • Takeaways & Limitations

    The authors intended the variant-enriched dataset to improve derived model performance on rare anatomical cases.

    Source: Background and Summary · p. 3

    Metastases and dislocated traumatic fractures were excluded from VerSe 2020.

    Source: Subject Selection · p. 6

Abstract

from arXiv · show

With the advent of deep learning algorithms, fully automated radiological image analysis is within reach. In spine imaging, several atlas- and shape-based as well as deep learning segmentation algorithms have been proposed, allowing for subsequent automated analysis of morphology and pathology. The first Large Scale Vertebrae Segmentation Challenge (VerSe 2019) showed that these perform well on normal anatomy, but fail in variants not frequently present in the training dataset. Building on that experience, we report on the largely increased VerSe 2020 dataset and results from the second iteration of the VerSe challenge (MICCAI 2020, Lima, Peru). VerSe 2020 comprises annotated spine computed tomography (CT) images from 300 subjects with 4142 fully visualized and annotated vertebrae, collected across multiple centres from four different scanner manufacturers, enriched with cases that exhibit anatomical variants such as enumeration abnormalities (n=77) and transitional vertebrae (n=161). Metadata includes vertebral labelling information, voxel-level segmentation masks obtained with a human-machine hybrid algorithm and anatomical ratings, to enable the development and benchmarking of robust and accurate segmentation algorithms.

Background and Summary

The VerSe 2020 dataset and challenge build on VerSe 2019 to address limited annotated CT spine data and the underrepresentation of rare anatomical variants. The paper reports an expanded, diverse dataset and challenge results supporting reliable automated spine segmentation.

  • VerSe 2020 expanded the dataset to 300 subjects, adding images from multiple institutions and four major scanner manufacturers, plus rare numerical and transitional vertebral variants.Source: Background and Summary · p. 3
  • Rare anatomical variants are underrepresented in unselected training datasets, and deep-learning algorithms fail in such cases.Source: Background and Summary · p. 3
  • The extended annotated dataset was intended to provide consistent, reliable ground truth for algorithm training and benchmarking.Source: Background and Summary · p. 4
  • The dataset was released through the second VerSe challenge at MICCAI 2020 and includes vertebral labels, voxel-level segmentations, and definitions of enumeration abnormalities and transitional vertebrae.Source: Background and Summary · p. 4
  • The authors describe VerSe 2020 as the largest publicly available CT spine imaging dataset with corresponding metadata for fully visualized vertebrae.Source: Background and Summary · p. 4
  • Successful 2019 and 2020 challenges confirm reliable fully automated deep-learning spine segmentation can be trained, and performance benefits from large, diverse datasets.Source: Background and Summary · p. 4

Methods

The dataset was assembled to broaden coverage of scanner vendors and spinal variants, then annotated through a semi-automated vertebral segmentation workflow with human review. Variant labeling used explicit anatomical criteria, while some sacral transitional vertebrae were excluded from segmentation and analysis.

Subject Selection

  • The dataset comprised 300 subjects: 150 multivendor cases and 150 anatomical-variant cases; 86 were from VerSe 2019 and 214 were new, with 36 VerSe 2019 variant subjects and 114 additionally selected subjects.Source: Subject Selection · pp. 5–6

Vertebral Segmentation

  • A semi-automated pipeline used CNN detection, Btrfly Net vertebral labeling, and U-Net segmentation before manual correction and neuroradiologist review.Corrections were limited to 1 mm accuracy; smaller-structure errors were not corrected.Source: Vertebral Segmentation · p. 7

Anatomical Variants

  • Thoracic versus lumbar labeling used rib length and alignment, with vertebral shape and facet orientation considered in ambiguous cases.Vertebrae with partial sacral fusion (Castellvi grades III and IV) were not segmented or included in further analysis.Source: Anatomical Variants · p. 8

Data Records

VerSe 2020 provides CT series, vertebral segmentations, labels, and metadata in public repositories. The records span multiple scanner vendors and include quantified anatomical variation.

Data Repositories and Storage

  • 319 CT image series from 300 subjects contain 4,142 vertebrae: 581 cervical, 2,255 thoracic, and 1,306 lumbar.The dataset and its training, public-test, and private-test subsets are publicly available under CC BY-SA 4.0.Source: Data Repositories and Storage · p. 9

Data Structure and File Formats

  • CT images and segmentation masks are stored in NIfTI format, while labels for all 4,142 segmented vertebrae are provided in JSON.The challenge data are divided into training, public-test, and private-test subsets of 100 subjects each.Source: Data Structure and File Formats · p. 9

CT Imaging and Scan Provenience

  • The dataset includes Philips, Siemens, Toshiba, and GE scans, with Philips accounting for 45.8% and Siemens for 32.3% of images.Heterogeneous, multivendor imaging was selected to support training and algorithm generalization; some data were not isotropic.Source: CT Imaging and Scan Provenience · p. 10

Anatomical Variants

  • Among scans depicting the lumbosacral region, 165 of 252 showed five lumbar vertebrae (65.5%), while 85 showed six (33.7%).The reported scan counts also include three scans with four lumbar vertebrae (1.2%).Source: Anatomical Variants · p. 10

Technical Validation

The validation reports localization and segmentation performance for the annotated data, alongside the human review process used to finalize masks.

Technical Validation

  • Voxel-level segmentation masks were generated with a human-machine hybrid algorithm, manually checked and corrected, then reviewed and approved by a neuroradiologist.Two neuroradiologists performed anatomical ratings in consensus.Source: Technical Validation · p. 11
  • The segmentation workflow achieved a best mean Dice score of 91.7% and localization error below 2 mm.The best-performing challenge Docker also achieved a mean vertebral identification rate of 95.6%.Source: Technical Validation · p. 11

Usage Notes

The dataset's metadata is intended to expand over time with additional anatomical and clinical measurements.

Usage Notes

  • Additional anatomical ratings, fracture grading, and bone mineral density measurements are slated for continuous updating and publication.The paper identifies the Open Science Framework as the publication site.Source: Usage Notes · p. 11

Code Availability

The VerSe-trained segmentation algorithm is publicly accessible, and the dataset and helper code are available online. Prior VerSe challenge submissions demonstrate the feasibility and potential of automated spine CT processing.

Code Availability

  • The VerSe-trained segmentation algorithm is publicly accessible, and the dataset, helper code, and evaluation scripts are available on GitHub.The helper code includes data reading, writing, and evaluation scripts.Source: Code Availability · p. 11
  • Submissions to the 2019 and 2020 VerSe challenges demonstrate the feasibility and potential of automated spine CT processing.The passage points to Sekuboyina et al. 19 for an overview of the submitted algorithms.Source: Code Availability · p. 11

Tables and Figures

The supplied figure and table notes identify manual versus automated processing steps and describe dataset classifications relevant to the displayed materials.

Tables and Figures

  • Unknown scans from a public dataset were included.The passage attributes the public dataset to reference 19.Source: Tables and Figures · p. 13
  • Lumbosacral vertebrae were graded using the Castellvi Classification.Source: Tables and Figures · p. 13
  • Figure 2 distinguishes fully automated processing steps from steps requiring manual interaction.Green boxes indicate automated steps; blue boxes indicate manual interaction.Source: Tables and Figures · p. 13
Loading 2103.06360v1…