Conditioned Variational Auto-encoder for Detecting Osteoporotic Vertebral Fractures

Detection of osteoporotic vertebral fractures in CT scans is a particularly challenging task that was never sufficiently addressed. This is due to the large variation among healthy vertebrae and the different shapes a fracture could present itself in. In this paper, we combine a reconstructing conditioned-variational auto-encoder architecture and a discriminating multi-layer-perceptron (MLP) to capture these different shapes. We also introduce a vertebrae-specific loss-weighing regime that maximizes the classification yield. Furthermore, we `look into' the learnt network by investigating the saliency maps, traversing the latent space and demonstrating its smoothness. Finally, we report our results on two datasets, including the publicly available xVertSeg dataset achieving an F1 score of 84%.