Efficient Multi-Label Panoptic Segmentation Strategy for Automated Fracture Detection in X-Ray Images

Pelvic ring fractures are severe injuries that require accurate anatomical reconstruction to ensure effective surgical treatment. Automated segmentation of pelvic structures in X-ray images can support computer-assisted preoperative planning and intraoperative navigation, improving accuracy and reducing procedure time. However, this task is particularly challenging due to complex anatomy, overlapping structures, and limited annotated datasets. This thesis presents the development of an efficient deep learning-based pipeline for automated multi-label panoptic segmentation of pelvic fractures in X-ray images, within the context of the PENGWIN 2024 Challenge. After reviewing the clinical background of pelvic fractures and the fundamentals of convolutional neural networks for medical image segmentation, a lightweight UNet-based architecture was implemented using PyTorch. Unlike previous multi-stage strategies requiring multiple models, the proposed method employs a single network with multiple output channels to simultaneously identify primary and secondary fracture fragments for each pelvic bone, reducing computational complexity. The model was trained and evaluated on a dataset of simulated fluoroscopic images derived from CT scans. Results show that the proposed approach achieves competitive segmentation performance while significantly improving efficiency and reducing inference time. These findings highlight the potential of compact deep learning models to support real-time clinical workflows in computer-assisted orthopaedic surgery.