Edge Guided Segmentation Mask Prediction for 3D CT Fracture Imagery

3D fractured femur segmentation is currently a subject of active investigation, particularly with machine learning approaches. However, challenges persist in scenarios where la-beled 3D segmentation maps, manually segmented by experts for training, are severely limited due to time consuming and expen-sive. We propose a new domain enriched approach for 3D femur segmentation through multi-task learning. Specifically, we intro-duce a novel 3D data-fitting branch to encourage the network to learn a comprehensive representation from 3D volumes. Inspired by the observation that 3D fracture structures project onto 2D image slices among the axial plane with informative and unique edge profiles, we propose a 2D architecture comprising a shared encoder and two decoders. These decoders respectively predict femur segmentation maps and edge profiles. Additionally, we develop a new support matching regularizer to enhance robust-ness via exploiting a segmentation mask-edge profile relationship. The middle feature maps from two branches are facilitated to capture and integrate information from both the 2D and 3D representations. Experiments performed on challenging 3D CT fracture datasets demonstrate that our method outperforms state- of-the-art alternatives on well-known evaluation metrics even when limited training data is available.