CNN-Based Osteoporotic Vertebral Fracture Prediction and Risk Assessment on MrOS CT Data: Impact of CNN Model Architecture

Osteoporosis is a metabolic disease causing structural degradation and increased fragility of bone. In particular, this affects the risk of vertebral fractures. Existing clinical methods for fracture risk assessment have been observed to have low sensitivities. Convolutional neural networks (CNNs) have shown promising capabilities in diagnostic and prognostic image analysis. We introduce a fully automatic pipeline for vertebral fracture risk assessment. It consists of a prognostic CNN which—based on vertebral body patches extracted from CT images—computes a sigmoid output score to classify whether the subject will experience a vertebral fracture within 10 years. The output scores of the vertebrae of a patient are averaged and then used in an age- and body mass index (BMI)-adjusted Cox proportional-hazards model to compute an individual vertebral fracture risk. We investigate the impact of choosing among different well-known CNN model architectures on prognostic classification performance and fracture risk assessment. We observed substantial variation of results across training epochs and data partitions, especially for larger model architectures. No systematic difference between 2D and 3D model versions was observed regarding standardized hazard ratio (sHR), while 2D models mostly performed similar or better than the corresponding 3D models regarding AUROC- and AUPRC-values and C-index. We achieved a C-index of 78% and a sHR of 2.6.