Abstract
In recent years, the number of pelvic fractures due to osteoporosis has been increasing with the aging of the population. In the case of fragility fractures of the pelvis (FFP) in the elderly, the deformed part is very small and not easy to detect on CT images, which may delay treatment after detection and prevent patients from recovering their functional prognosis. Therefore, it desires computer-aided diagnosis (CAD) system that can automatically detect FFP from CT images to improve the diagnosis ability of physicians and reduce their workload. Conventional methods are based on two-dimensional image analysis of simple X-ray images or CT images, making it difficult to detect small fragility fractures distributed in three dimensions. We proposed a new method called boring survey based fracture detection (BSFD), which surveys fractures from the bone surface to the internal bone. In the BSFD method, a bone surface equivalence plane is obtained from 3D CT images, a 3D feature vector consisting of CT values is assigned to each point on the equivalence plane, and the fracture probability is estimated at each point by a trained 3D convolutional neural network (CNN) model. The CNN is trained using the fracture probability, which is calculated from the 3D Chamfer distance from the annotated fracture region. We pre-train several CNN models with medical 3D data, and train them again with our data.
