Single-Shot X-ray to Multi-View Projections for 3D Pork Shoulder Bone Analysis

Pork is an important meat product for the European Union, which exported over 4.2 million tons in 2023, valued at €8.1 billion. Automating the labor-intensive deboning process is of significant interest, particularly through the development of advanced inline inspection systems capable of analyzing pork shoulder bone structures. While computed tomography (CT) systems provide high-contrast 3D reconstructions, their large size and high-cost present substantial barriers to adoption in industrial meat processing. This study addresses these challenges by introducing a novel approach that uses a single X-ray projection in combination with deep neural networks to predict the 3D segmentation map of pork shoulder bone structures using conventional reconstruction algorithms. To this end, U-Net neural network variants were trained on high-resolution CT scans of 90 pork shoulders. These scans were augmented with synthetic data to simulate different orientations on a conveyor belt, ensuring the model’s robustness. The minimum number of X-ray projections needed for accurate reconstruction was determined based on simulations, and 60 evenly spaced projections between 0° and 180° were found optimal. The Feldkamp-Davis-Kress (FDK) algorithm was chosen for its efficiency and cost-effectiveness in inline processing. The model achieved a Dice score of 0.94 and an SSIM of 0.96 on test data, demonstrating its ability to predict 59 missing projections and reconstruct the 3D bone structure accurately. The method that is proposed in this paper has the potential to advance meat processing by enhancing deboning precision, reducing waste, and streamlining operations.