Bone Quality Classification of Dual Energy X-ray Absorptiometry Images using Convolutional Neural Network Models

Abstract

The assessment of bone trabecular quality degradation is important for the detection of diseases such as osteoporosis. The gold standard for its diagnosis is the Dual Energy X-ray Absorptiometry (DXA) image modality. The analysis of these images is a topic of growing interest, especially with artificial intelligence techniques. This work proposes the detection of a degraded bone structure from DXA images using some approaches based on the learning of Trabecular Bone Score (TBS) ranges. The proposed models are supported by intelligent systems based on convolutional neural networks using two kinds of approaches: ad hoc architectures and knowledge transfer systems in deep network architectures, such as AlexNet, ResNet, VGG, SqueezeNet, and DenseNet retrained with DXA images. For both approaches, experimental studies were made comparing the proposed models in terms of effectiveness and training time, achieving an F1-Score result of approximately 0.75 to classify the bone structure as degraded or normal according to its TBS range.