OP0301 PREDICTION OF LOW BONE MINERAL DENSITY AND FRAX SCORE BY ASSESSING HIP BONE TEXTURE WITH DEEP LEARNING

Abstract

Background:Osteoporosis is a widespread health concern associated with an increased risk of fractures in individuals with low bone mineral density (BMD). Dual-energy x-ray absorptiometry (DXA) is the gold standard to measure BMD, but methods based on the assessment of plain films, such as the digital radiogrammetry,1are also available. We describe a novel approach based on the assessment of hip texture with deep learning to estimate BMD.Objectives:To compare the BMD estimated by assessing hip texture using a deep learning model and that measured by DXA.Methods:In this study, we identified 1,203 patients who underwent DXA of left hip and hip plain film within six months. The dataset was split into a training set with 1,024 patients and a testing set with 179 patients. Hip images were obtained and regions of interest (ROI) around left hips were segmented using a tool based on the curve Graph Convolutional Network. The ROIs are processed using a Deep Texture Encoding Network (Deep-TEN) model,2which comprises the first 3 blocks of Residual Network with 18 layers (ResNet-18) model followed by a dictionary encoding operator (Figure 1). The encoded features are processed using a fully connected layer to estimate BMD. Five-fold cross-validation was conducted. Pearson’s correlation coefficient was used to assess the correlation between predicted and reference BMD. We also test the performance of the model to identify osteoporosis (T-score ≤ -2.5)Figure 1.Schematic representation of deep learning models to extract and encode texture features for estimation of hip bone density.Results:We included 151 women and 18 men in the testing dataset (mean age, 66.1 ± 1.7 years). The mean predicted BMD was 0.724 g/cm2compared with the mean BMD measured by DXA of 0.725 g/cm2(p = 0.51). Pearson’s correlation coefficient between predicted and true BMD was 0.88. The performance of the model to detect osteoporosis/osteopenia was shown in Table 1. The positive predictive value was 87.46% for a T-score ≤ -1 and 83.3% for a T-score ≤ -2.5. Furthermore, the mean FRAX® 10-year major fracture risk did not differ significantly between scores based on predicted (6.86%) and measured BMD (7.67%, p=0.52). The 10-year probability of hip fracture was lower in the predicted score (1.79%) than the measured score (2.43%, p = 0.01).Table 1.Performance matrices of the deep texture model to detect osteoporosis/osteopeniaT-score ≤ -1T-score ≤ -2.5Sensitivity91.11%(95% CI, 83.23% to 96.08%)33.33%(95% CI, 17.29% to 52.81%)Specificity86.08%(95% CI, 76.45% to 92.84%)98.56%(95% CI, 94.90% to 99.83%)Positive predictive value88.17%(95% CI, 81.10% to 92.83%)83.33%(95% CI, 53.58% to 95.59%)Negative predictive value89.47%(95% CI, 81.35% to 94.31%)87.26%(95% CI, 84.16% to 89.83%)Conclusion:This study demonstrates the potential of the bone texture model to detect osteoporosis and to predict the FRAX score using plain hip radiographs.