Fast quantitative bone marrow lesion measurement on knee MRI for the assessment of osteoarthritis

Frank Preiswerk; Meera S Sury; Jeremy R Wortman; Gesa Neumann; William Wells; Jeffrey Duryea

doi:10.1016/j.ocarto.2022.100234

Fast quantitative bone marrow lesion measurement on knee MRI for the assessment of osteoarthritis

Osteoarthr Cartil Open. 2022 Jan 10;4(1):100234. doi: 10.1016/j.ocarto.2022.100234. eCollection 2022 Mar.

Authors

Frank Preiswerk¹, Meera S Sury¹, Jeremy R Wortman², Gesa Neumann³, William Wells¹, Jeffrey Duryea¹

Affiliations

¹ Department of Radiology, Brigham and Women's Hospital, Harvard Medical School, Boston, MA, 02115, USA.
² Department of Radiology, Lahey Hospital and Medical Center, Tufts University School of Medicine, Burlington, MA, 01805, USA.
³ Department of Radiology, Boston Medical Center, Boston University, Boston, MA, 02118, USA.

Abstract

Objective: Knee osteoarthritis (KOA) is a prevalent disease with a high economic and social cost. Magnetic resonance imaging (MRI) can be used to visualize many KOA-related structures including bone marrow lesions (BMLs), which are associated with OA pain. Several semi-automated software methods have been developed to segment BMLs, using manual, labor-intensive methods, which can be costly for large clinical trials and other studies of KOA. The goal of our study was to develop and validate a more efficient method to quantify BML volume on knee MRI scans.

Materials and methods: We have applied a deep learning approach using a patch-based convolutional neural network (CNN) which was trained using 673 MRI data sets and the segmented BML masks obtained from a trained reader. Given the location of a BML provided by the reader, the network performed a fully automated segmentation of the BML, removing the need for tedious manual delineation. Accuracy was quantified using the Pearson's correlation coefficient, by a comparison to a second expert reader, and using the Dice Similarity Score (DSC).

Results: The Pearson's R² value was 0.94 and we found similar agreement when comparing two readers (R² = 0.85) and each reader versus the DL model (R² = 0.95 and R² = 0.81). The average DSC was 0.70.

Conclusions: We developed and validated a deep learning-based method to segment BMLs on knee MRI data sets. This has the potential to be a valuable tool for future large studies of KOA.

Keywords: Bone marrow lesions; Knee osteoarthritis; MRI; Software assessment.