Predicting rat lumbar vertebral failure patterns as synthetic μCT images using a deep convolutional generative adversarial network

Abstract

This study aims to develop a generative deep learning (DL) model to predict fracture patterns through creation of synthetic 3D μCT images of fractured rat lumbar vertebrae. The proposed model is a 3D conditional generative adversarial network (cGAN). The dataset contained sequential 3D μCT images of rat lumbar vertebrae acquired during axial compressive loading (unloaded, after 1500 μm of displacement, after fracture). Three experiments were run: unloaded input predicting 1500 μm, 1500 μm predicting fracture, and unloaded predicting fracture. The cGAN was trained on 64 μCT images of rat lumbar vertebral motion segments and was validated on 8 images. Quantitative metrics (dice similarity coefficient (DSC), Jaccard index (JAC), Fréchet inception distance (FID), structural similarity index measure (SSIM)) assessed predicted image quality. Qualitative measures investigating fracture location and disease severity (or lack thereof) were also assessed. The unloaded to 1500 μm experiment generated realistic examples of loaded (unfractured) rat vertebrae. These included maintenance of the presence of metastatic disease when relevant and yielded the best quantitative metrics. The 1500 μm to fracture experiment performed significantly better on the FID and SSIM metrics than the unloaded to fracture configuration. The 1500 μm to fracture experiment predicted more true positive fractures and fewer false negative fractures than the unloaded to fracture experiment. Both fracture experiments had a low false positive fracture prediction rate (<10 %). The presented cGAN generates realistic rat lumbar vertebrae failure patterns as 3D μCT images and shows promise for future generative DL applications to model damage behaviour of biological structures.