gms | German Medical Science

German Congress of Orthopaedics and Traumatology (DKOU 2019)

22. - 25.10.2019, Berlin

Detecting bone marrow edemas from plain radiographs using deep learning: data from the osteoarthritis initiative (OAI)

Meeting Abstract

  • presenting/speaker Tiago Paixao - ImageBiopsy Lab, Vienna, Austria
  • Christoph Götz - ImageBiopsy Lab, Vienna, Austria
  • Richard Ljuhar - ImageBiopsy Lab, Vienna, Austria
  • Hans-Peter Dimai - Medizinische Universtät Graz, Graz, Austria
  • Stefan Nehrer - Zentrum für Regenerative Medizin und Orthopädie, Donau Universität Krems, Krems, Austria

Deutscher Kongress für Orthopädie und Unfallchirurgie (DKOU 2019). Berlin, 22.-25.10.2019. Düsseldorf: German Medical Science GMS Publishing House; 2019. DocAB40-727

doi: 10.3205/19dkou321, urn:nbn:de:0183-19dkou3219

Published: October 22, 2019

© 2019 Paixao et al.
This is an Open Access article distributed under the terms of the Creative Commons Attribution 4.0 License. See license information at



Objectives: Bone marrow edemas are defined as localized build-up of fluids within the bone marrow, often leading to swelling, which are strongly associated with pain and progression of joint deterioration. Bone marrow lesions are currently only able to be diagnosed through magnetic resonance imaging. However, bone marrow lesions have been link to altered bone density, raising the hypothesis that these alterations can be detected on plain radiographs. Here we ask whether a convolutional neural network is able to detect bone marrow lesions from plain radiographs of the knee.

Methods: We collected radiographic images and readings from the data coming from the large longitudinal study focused on knee osteoarthritis: the Osteoarthritis Initiative (OAI) study. Each individual in the OAI study was imaged in both radiography and MRI modalities at a number of timepoints. MRI readings according to the MRI Osteoarthritis Knee Scoring method (MOAKS) were provided by the OAI study (project 22). We matched radiographic images to bone marrow lesion readings (MOAKS BML subscore) from magnetic resonance imaging to obtain a dataset of 5214 radiographic images of single knees, labeled with respect to the presence, number, and location of bone marrow lesions in the corresponding MRI. We split this dataset into train (2085), validation (1564), and testing (1565) sub-datasets which were used to train, and independently cross-validate, an 8 layer convolutional neural network to detect the presence of bone marrow lesions from plain radiographs. Class imbalance was dealt with by oversampling the minority class.

Results and conclusion: We find that our convolutional neural network achieves an average weighted class accuracy greater than 70% in the test dataset, with a sensitivity of 0.72 and a specificity of 0.69, when classifying presence/absence of bone marrow lesions from plain radiographs. Furthermore, saliency analysis indicates that the neural network can discriminate the joint compartment in which bone marrow lesions are located.

Our results show that it is possible to detect bone marrow lesions from plain radiographs. This is not completely unexpected since bone marrow lesions have been associated with changes in bone density which could potentially be detectable on a plain radiograph. These results suggest that it is possible to acquire indications for the presence of bone marrow lesions from radiographs, which could lead to a change in clinical management. Furthermore, early identification of bone marrow lesions from plain radiographs has the promise to further elucidate the role of these lesions on the etiology of knee osteoarthritis.