A Vertebral Segmentation Dataset with Fracture Grading

Maximilian T. Löffler, Anjany Sekuboyina, Alina Jacob, Anna-Lena Grau, Andreas Scharr, Malek El Husseini, Mareike Kallweit, Claus Zimmer, Thomas Baum, Jan S. KirschkeRadiology 2020 · DOI

TL;DR

This paper introduces a novel, publicly available dataset of 160 CT scans with detailed vertebral segmentation masks and fracture gradings. This dataset aims to address the critical underdiagnosis of osteoporotic vertebral fractures by providing a robust resource for training and validating AI models for automated spine analysis.

What Problem Does This Paper Address?

Accurate identification and segmentation of vertebrae are fundamental steps for assessing spinal pathologies, including fractures. However, vertebral fractures, particularly osteoporotic ones, are frequently underdiagnosed, with up to 85% missed on CT scans. This underdiagnosis poses a significant challenge, especially given the increasing workload of radiologists. While deep learning shows promise for automated diagnosis, it requires large, well-annotated datasets. Existing public datasets for vertebral segmentation are limited in size and scope, often lacking comprehensive annotations for fractures or including limited vertebral regions. This paper addresses this gap by creating a large-scale, annotated dataset specifically designed to facilitate the development of AI tools for automated vertebral segmentation and fracture grading.

Methods

The authors employed a three-step approach for vertebral segmentation. First, CT data were anonymized and preprocessed to a standardized resolution. Second, a deep learning framework utilizing a fully convolutional network (CNN) was implemented. This framework incorporated a CNN for initial spine detection, a Btrfly Net for labeling vertebrae on maximum intensity projections, and an improved U-Net for detailed segmentation of vertebral patches. The U-Net was trained on existing public datasets and then retrained with the finalized segmentations from their new dataset. Finally, the segmentation masks underwent manual refinement by trained medical students and neuroradiologists, with specific exclusion of non-physiological material like implants. Fracture assessment involved a semiquantitative grading method by Genant et al., classifying fractures based on height loss and type. Opportunistic measurements of lumbar bone mineral density (BMD) were also performed.

Datasets

The paper introduces a new, publicly available dataset comprising 160 CT image series from 141 patients. This dataset is split into training (80 series), public validation (40 series), and secret test (40 series) subsets. It includes segmentation masks for 1725 fully visualized vertebrae and annotations for vertebral fractures and foreign material. The dataset also contains opportunistic BMD measurements for the lumbar spine. This dataset is intended to be a benchmark for vertebral labeling and segmentation challenges, building upon and significantly expanding existing smaller datasets like the Computational Spine Imaging 2014 Workshop dataset and the xVertSeg challenge dataset.

Key Insights and Ideas

  • The critical need for large-scale, annotated data: The paper strongly emphasizes that the underdiagnosis of vertebral fractures is a significant clinical problem, and the lack of comprehensive datasets is a major bottleneck for developing effective AI solutions.
  • Deep learning's potential for automated spine analysis: The authors highlight the success of deep learning, particularly CNNs, in various medical imaging tasks and propose their dataset as a catalyst for advancing automated spine segmentation and fracture detection.
  • Comprehensive annotation is key: Beyond just segmentation, the inclusion of fracture grading and BMD measurements makes this dataset particularly valuable for developing AI that can provide clinically relevant information.
  • The challenge of manual annotation: The paper implicitly acknowledges the significant effort involved in creating such a dataset, detailing the multi-stage refinement process, which underscores the value of the final product.

Key Results

The dataset contains 160 CT image series and segmentation masks for 1725 vertebrae. It includes annotations for vertebral fractures, with 289 identified fractures across different grades. The dataset also provides BMD measurements for 141 patients. The authors report that their dataset represents a more than fourfold increase in annotated data compared to previously available datasets, particularly for pathological and cervical vertebrae.

Limitations and Future Work

The authors acknowledge several limitations:

  • Age restriction: The dataset only includes patients older than 30 years, potentially limiting the reliability of trained algorithms for younger individuals.
  • Exclusion of certain pathologies: Bone metastasis and primary bone tumors were excluded, and the dataset does not cover all normal variants or vertebral abnormalities.
  • Focus on edge-enhancing reconstructions: While practical for current interpretation, future work could explore soft-tissue kernels and iterative reconstruction algorithms.
  • Retrospective design and resolution: The retrospective nature means isotropic resolution was not always available, and a 1-mm spatial resolution limit was imposed for manageability.
  • Single-rater annotation: While refined, the final segmentation approval was by a single rater, introducing potential variability.

Future work suggested includes:

  • MRI segmentation: Extending automated segmentation algorithms to MRI data.
  • Multi-rater fusion: Investigating methods to incorporate variability from multiple raters.
  • Addressing degenerative changes: Developing methods to handle complex cases with degenerative changes that can obscure vertebral borders.
  • Clinical benefit demonstration: Proving that computer-aided diagnosis using such datasets can indeed benefit patients.

Summary updated

Abstract

Published under a CC BY 4.0 license. <i>Supplemental material is available for this article</i>.
BibTeX
@inproceedings{loffler2020vertebral,
  author = {Maximilian T. Löffler and Anjany Sekuboyina and Alina Jacob and Anna-Lena Grau and Andreas Scharr and Malek El Husseini and Mareike Kallweit and Claus Zimmer and Thomas Baum and Jan S. Kirschke},
  title = {{A Vertebral Segmentation Dataset with Fracture Grading}},
  booktitle = {Radiology},
  year = {2020},
  doi = {10.1148/ryai.2020190138},
}
Figure