About Dataset

Introduction

Osteoarthritis (OA) is a degenerative disease of the joints and manifests enormous societal and personal costs. Over time, the "wear-and-tear" of protective tissues, such as the articular cartilage in the knee, can be lead to pain, immobility, and joint disfunction. The morphology or shape of these tissues are of high interest for clinicians and researchers, and the use of high resolution and 3D imaging techniques like magnetic resonance imaging (MRI) have allowed for the study of the relationship between tissue structure and disease onset and progression.

This dataset is an ensemble of two efforts to automatically segment various tissues from the Osteoarthritis Initiative (OAI), a NIH-funded observational study of knee OA. Additional information about the OAI can be found here: https://nda.nih.gov/oai/. In this dataset, you will find tissue segmentations of 46,824 knee MRIs acquired from 4,795 subjects across the OAI imaging visits.

Please cite the papers below if you use this dataset in your research.

Deep Learning-based Segmentation

Deep convolutional neural networks were trained to segment bone (femur, tibia, and patella) [1][2], cartilage (femoral, tibial, and patellar cartilage) and meniscus [3] from DESS-we MRI. Details regarding image labeling, model training, and evaluation can be found in the original texts. The performance of the models in unseen test sets is summarized below for convenience:

All available DESS-we MRIs in the OAI were inferred using the two sets of models. Ensembling of the bone and cartilage/meniscus masks was performed element-wise. For any conflicts between model results, cartilage and meniscus were prioritized over bone.

How to Access Data

This dataset is partitioned into directories of each imaging visit of the OAI schedule. The naming convention of each file is designated as XXXXXXX_YYYY_ZZ.npz, where X = subject ID, Y = left or right knee, and Z = visit code.

The segmentation data is stored in a compressed numpy file format. Below is an example of how to load a singular mask.

import numpy as np
f = np.load('path/to/file.npz')
mask = f['x']

Citations

[1] Morales Martinez A, Caliva F, Flament I, et al. Learning osteoarthritis imaging biomarkers from bone surface spherical encoding. Magn Reson Med 2020;84:2190–203.

[2] Calivá F, Kamat S, Morales Martinez A, Majumdar S, Pedoia V. Surface spherical encoding and contrastive learning for virtual bone shape aging. Med Image Anal. 2022 Apr;77:102388.

[3] Iriondo C, Liu F, Calivà F, et al. Towards understanding mechanistic subgroups of osteoarthritis: 8-year cartilage thickness trajectory analysis. J Orthop Res 2021;39:1305–17.