Source-linked AI summary
A large annotated medical image dataset for the development and evaluation of segmentation algorithms
Amber L. Simpson, Michela Antonelli, Spyridon Bakas, Michel Bilello, Keyvan Farahani, Bram van Ginneken, Annette Kopp-Schneider, Bennett A. Landman, Geert Litjens, Bjoern Menze, Olaf Ronneberger, Ronald M. Summers, Patrick Bilic, Patrick F. Christ, Richard K. G. Do, Marc Gollub, Jennifer Golia-Pernicka, Stephan H. Heckers, William R. Jarnagin, Maureen K. McHugo, Sandy Napel, Eugene Vorontsov, Lena Maier-Hein, M. Jorge Cardoso
TL;DR
Medical image segmentation needs high-quality expert-labeled data, yet sharing and benchmarking remain constrained by privacy barriers and small evaluations. The paper constructs and standardizes a large, multi-institutional collection of ten annotated datasets spanning variable tasks and makes it openly reusable. The resulting resource supports comprehensive benchmarking and broad access to medical image data, while its retrospective clinical acquisition introduces substantial protocol variation.
Problem
High-quality labeled medical imaging data are needed for semantic segmentation, but privacy concerns restrict sharing and many algorithms are evaluated on small samples with limited comparisons.
Method
The authors assembled ten curated annotated datasets from multiple institutions, de-identified and reformatted the images, and made the collection openly available for reuse.
Results
The collection provides 2,633 three-dimensional images spanning multiple anatomies, modalities, institutions, and real-world clinical applications.
Takeaways & Limitations
The resource enables objective benchmarking of general-purpose segmentation methods and open access to medical image data for researchers.
Takeaways & Limitations
Because the data were acquired retrospectively through routine clinical scanning, acquisition and reconstruction parameters vary substantially within and across institutions.
Abstract
from arXiv · showhide
Semantic segmentation of medical images aims to associate a pixel with a label in a medical image without human initialization. The success of semantic segmentation algorithms is contingent on the availability of high-quality imaging data with corresponding labels provided by experts. We sought to create a large collection of annotated medical image datasets of various clinically relevant anatomies available under open source license to facilitate the development of semantic segmentation algorithms. Such a resource would allow: 1) objective assessment of general-purpose segmentation methods through comprehensive benchmarking and 2) open and free access to medical image data for any researcher interested in the problem domain. Through a multi-institutional effort, we generated a large, curated dataset representative of several highly variable segmentation tasks that was used in a crowd-sourced challenge - the Medical Segmentation Decathlon held during the 2018 Medical Image Computing and Computer Aided Interventions Conference in Granada, Spain. Here, we describe these ten labeled image datasets so that these data may be effectively reused by the research community.
Background & Summary
Medical image segmentation research is constrained by limited access to high-quality labeled data and small-scale evaluations. The paper addresses this gap by creating a large, open, multi-institutional collection spanning variable segmentation tasks for benchmarking and reuse.
- High-quality labeled medical imaging data are essential for developing semantic segmentation algorithms, but institutions face privacy-related barriers to sharing them.Removing protected health information can be expensive and time-consuming.
- Many reported algorithms are validated on few samples and compared with few existing approaches, limiting assessment of generalizability and incremental progress.
- The authors created a large, open-source, manually annotated dataset covering multiple anatomical sites and highly variable segmentation tasks.The effort was conducted through a multi-institutional collaboration.
- The collection was designed to support objective benchmarking of general-purpose segmentation methods and broaden access to medical image data.
Methods
The Medical Segmentation Decathlon assembled and standardized ten heterogeneous medical imaging datasets from multiple institutions, anatomies, modalities, and clinical settings. The datasets include expert or semi-automatic annotations and were reformatted to improve interoperability and accessibility.
- Dataset collection: 2,633 three-dimensional images were collected across multiple anatomies, modalities, and institutions representative of real-world clinical applications.All images were de-identified and reformatted to reduce reliance on specialized medical-imaging software.
- Dataset collection: The ten datasets were selected for availability and suitability for semantic segmentation algorithm development.
- Annotations: Annotations were produced through expert manual tracing, expert segmentation, or semi-automatic tools depending on the dataset.The hippocampus was manually traced, while spleen segmentation used a semi-automatic workflow and colon tumours were segmented by an expert radiologist.
- Data processing: All images were converted from DICOM to NIfTI, an open format intended to ensure consistency, interoperability, and use without proprietary software.
Competing financial interests
The authors disclose research funding, consultancy, royalties, grants, and research support involving pharmaceutical, imaging, and technology organizations.
- G.L. reports funding from Philips Digital Pathology Solutions, a consultancy role for Novartis, and grants from Dutch and Dutch research organizations.
- R.M.S. reports royalties from iCAD, Philips, ScanMed, and PingAn, plus research support from PingAn and NVIDIA.
Data Citations
The paper identifies source datasets and provides visual and structural information supporting reuse of the downloadable collection. Dataset examples show label-color conventions, while the download includes machine-readable metadata.
- The paper cites source collections for glioblastoma, low-grade glioma, and related segmentation labels and radiomic features.
- Figure 1 presents exemplar images and labels for each dataset, using blue, white, and red for labels 1, 2, and 3 when applicable.Not all tasks contain three labels.
- The download includes a JSON descriptor alongside image and label directories for each dataset.
- The collection’s Table 1 identifies medical image datasets available for download and reuse.