An Informative Machine-Learning Tool for Diagnosis of Osteoporosis using Routine Femoral Neck Radiographs

Abstract

Aim/Purpose: The aim of the study was to analyze the structure of the bone tissue by using texture analysis of the bone trabeculae, as visualized in a routine radiograph of the proximal femur . This could provide objective information regarding both the mineral content and the spatial structure of bone tissue. Therefore, machine-learning tools were applied to explore the use of texture analysis for obtaining information on the bone strength. Background: One in three women in the world develops osteoporosis, which weakens the bones, causes atraumatic fractures and lowers the quality of life. The damage to the bones can be minimized by early diagnosis of the disease and preventive treatment, including appropriate nutrition, bone-building exercise and medications. Osteoporosis is currently diagnosed primarily by DEXA (Dual Energy X-ray Absorptiometry), which measures the bone mineral density alone. However, bone strength is determined not only by its mineral density but also by the spatial structure of bone trabeculae. In order to obtain valuable information regarding the bone strength, the mineral content and the spatial structure of the bone tissue should be objectively assessed. Methodology: The study includes 17 radiographs of in-vitro femurs without soft tissue and 44 routine proximal femur radiographs (15 subjects with osteoporotic fractures and 29 without a fracture). The critical force required to fracture the in-vitro femurs was measured and the bones were divided into two groups: 11 solid bones with critical fracture force higher than 4.9kN and 6 fragile bones with critical fracture force lower than 4.9kN. All the radiographs included an aluminum step-wedge for calibrating the gray-levels values (See Figure 3). An algorithm was developed to automatically adjust the gray levels in order to yield equal brightness and contrast. Findings: The algorithm characterized the in-vitro bones with as fragile or solid with an accuracy of 88%. For the radiographs of the patients, the algorithm characterized the bones as osteoporotic or non-osteoporotic with an accuracy of 86%. The most prominent features for estimating the bone strength were the mean gray-level, which is related to bone density, and the smoothness, uniformity and entropy, which are related to the spatial distribution of the bone trabeculae. Impact on Society: Analysis of bone tissue structure, using machine-learning tools will provide a significant information on the bone strength, for the early diagnosis of osteoporosis. The structure analysis can be performed on routine radiographs of the proximal femur, with high accuracy. Future Research: The algorithm for automatic structure analysis of bone tissue as visualized on a routine femoral radiograph should be further trained on a larger dataset of routine radiographs in order to improve the accuracy of assessing the bone strength.