SUN-814 Artificial Intelligence Image Analysis Of Osteoporosis In Dual-energy X-ray Absorptiometry - Preset Data

Abstract Disclosure: Y. Chung: None. W. Kim: None. Deep learning techniques have recently emerged as a powerful method for analyzing dual-energy X-ray absorptiometry (DXA) images of spinal bones. These techniques utilize neural networks to differentiate between normal and osteoporotic bone patterns, enabling automated classification of patient health status. While conventional radiography remains widely used in clinical practice, its application for osteoporosis assessment has historically been limited by qualitative interpretation and subjective analysis. Recent advances in machine learning (ML), however, have enabled the transformation of standard DXA imaging into a quantitative diagnostic tool. This study developed an ML-based diagnostic protocol using Keras-based image analysis combined with a k-nearest neighbors (KNN) algorithm. Our approach integrated two key components: (1) preprocessed randomized osteoporosis images and (2) the average standard deviation of lumbar spine distances from DXA scans, including those without obvious symptoms. This dual-method analysis correlated radiographic features with clinical data to improve diagnostic accuracy. The study employed randomized normal and osteoporotic images for model training. Initial analysis used a Keras model, followed by implementation of TensorFlow with a VGG-16 convolutional neural network (16-layer depth) from the Python library. We further developed a customized model incorporating an image directory and an average standard deviation equation, implemented in Python. Results demonstrated that pretraining with well-defined, randomized spine images significantly enhanced diagnostic performance when applied to real DXA scans, achieving 95.2% accuracy. In contrast, the KNN algorithm alone showed limited capability for osteoporosis identification, reaching only 83.53% accuracy even with clearly distinguishable images. This finding suggests that KNN alone is inadequate for reliable osteoporosis detection. Image analysis revealed distinct morphological differences: normal samples exhibited 18.63% bone occupancy, while osteoporotic samples showed 25.49%. Spatial distribution mapping of bone structures demonstrated that osteoporotic cases had 37.67% average empty space compared to 13.44% in healthy individuals. Measurement variability was significantly higher in osteoporosis cases (standard deviation: 89.2 vs. 33.6 in healthy controls). In conclusion, while KNN proves insufficient for raw DXA image analysis (83.53% accuracy), the VGG-16 architecture demonstrated superior diagnostic capability, achieving 95.2% accuracy in osteoporosis detection. The 16-layer convolutional network's deep feature extraction capabilities proved particularly effective in identifying subtle osteoporotic patterns that eluded traditional machine learning approaches. Presentation: Sunday, July 13, 2025