Deep learning-based automated spine fracture type identification with Clinically validated GAN generated CT images

Automatic type identification of sub-axial spine fractures is of prime importance for orthopaedicians to reduce image interpretation time and increase patient care time. But identifying fracture types is challenging due to imbalanced datasets. In this work, CT scan images of fractured spine has been collected from a Tertiary Care hospital and extended Deep Convolutional Generative Adversarial Network (DCGAN) architecture is developed for generating spine fracture images that overcomes the imbalanced dataset problem. These enhanced dataset are clinically evaluated with Two Visual Turing Tests (VTTs): the first test to "identify real and generated images" and second test to determine "type of fractures in the generated images." The first VTT demonstrates that generated images of fractures are realistic and that even spine surgeons have difficulty in distinguishing them from real. The second VTT demonstrates that fracture lines are clearly visible in the generated images. The VTT results are analyzed using Fleiss Kappa statistical techniques to determine the inter-observer reliability of spine surgeons' clinical evaluation. The results showed high interobserver agreement for "type identification" in the generated images. The clinically evaluated generated images are provided to the proposed ensemble based type identification model, which outperformed other models in type identification.