COMPARISON BETWEEN BML VOLUMES FROM A SEMI-AUTOMATED SOFTWARE METHOD AND BML VOLUMES FROM A DEEP LEARNING ALGORITHM: DATA FROM THE OSTEOARTHRITIS INITIATIVE (OAI)

Abstract

MRI can be used to assess several structural changes related to knee OA (KOA) including bone marrow lesions (BMLs), which are important structural biomarkers linked to pain and disease progression. Software image analysis methods have been shown to provide objective quantitative measurements of BML volume, and deep learning (DL) algorithms offer the potential to eliminate or reduce the reader time substantially. The goal of our study was to compare BML volumes determined by a deep learning (DL) algorithm to those measured by a well-validated semi-automated (SA) approach and, ultimately, demonstrate that the DL algorithm can be used in place of the SA method. We measured BML volumes in index knees of 700 subjects from the OAI at the BL, 12mo and 24mo time points, for a total of 2,084 knees, using the SA and DL methods. The DL method used a U-Net convolutional neural network (CNN) with 5 convolution levels and was trained using Dice coefficients. Each knee was subdivided into several sub regions: femur (F), tibia (T), patella (P), medial femur, tibia, and patella (MF, MT, MP) lateral femur, tibia, and patella (LF, LT, and LP), the trochlea (TR) and a measure of the patellofemoral (PF) subregion defined as the combination of P and TR. A least-squared fit was performed comparing the DL and SA for the total knee BML volume and each of the sub regions. Figure 1 shows scatter and Bland Altmam plots of the DL versus SA total knee volume, demonstrating a strong correlation but an apparent bias as the DL method consistently underestimates the SA volume for larger values. The table provides the slopes and R 2 values for total knee and sub regions. While there was a strong association between the DL and SA methods, a consistent bias was evident across all sub regions in the deviation of the slopes from a value of 1. The reason for the bias appears to be the diffuse nature of BMLs on MRI and the fact that the DL algorithm does not capture all such voxels. Whether an alternative to training with Dice, or the application of a correction factor can be applied merits further investigation. The results suggest that the DL approach may be useful for future studies that require processing of a large number of MRI scans. NIH AR071409 DICLOSURE STATEMENT ACKNOWLEDGMENT CORRESPONDENCE ADDRESS: jduryea@bwh.harvard.edu