Uploaded June 2025 | Updated September 2026, 10 hours ago
This is a video about Tim Schär's bachelor's thesis at the Bern University of Applied Sciences in Switzerland. Adolescent idiopathic scoliosis patients require frequent imaging during growth periods; however, current clinical approaches present significant limitations. While CT scans provide excellent bone contrast for segmentation, they expose patients to substantial ionizing radiation. Biplanar EOS X-ray imaging offers reduced radiation exposure while capturing anatomically relevant standing-position deformities. However, automated segmentation of overlapping rib structures from these 2D projections remains challenging. The severe scarcity of labeled EOS datasets further hinders the development of machine learning based segmentation models. This thesis develops a comprehensive automated pipeline that transforms biplanar EOS images into patient-specific 3D rib cage models through domain adversarial training and geometric reconstruction. To address data limitations, synthetic training datasets are generated from over 750 CT volumes using physics-based Digitally Reconstructed Radiography with multilabel segmentation preservation. A custom domain adversarial neural network architecture employs a twelve-channel output to maintain individual rib continuity while reducing the domain gap between synthetic and clinical images through gradient reversal mechanisms. The final component reconstructs volumetric
models by establishing a mathematical correspondence between orthogonal projections, achieved through centerline extraction, and parametric representation.
The segmentation network achieved Dice scores of 0.95 for coronal and 0.93 for sagittal views on synthetic data, with domain adversarial training producing measurable improvements in generalization to clinical EOS images. The 3D reconstruction component successfully captures patient-specific anatomical characteristics, including the overall shape of the rib cage, individual rib curvature patterns, the spatial relationships between rib pairs, and patient-specific spinal deformations with associated rib adaptations. This work establishes a foundation for extracting comprehensive anatomical information from limited projection data while providing methodological frameworks for medical imaging challenges involving synthetic data generation and cross-domain adaptation.
This is a video about Tim Schär's bachelor's thesis at the Bern University of Applied Sciences in Switzerland. Adolescent idiopathic scoliosis patients require frequent imaging during growth periods; however, current clinical approaches present significant limitations. While CT scans provide excellent bone contrast for segmentation, they expose patients to substantial ionizing radiation. Biplanar EOS X-ray imaging offers reduced radiation exposure while capturing anatomically relevant standing-position deformities. However, automated segmentation of overlapping rib structures from these 2D projections remains challenging. The severe scarcity of labeled EOS datasets further hinders the development of machine learning based segmentation models. This thesis develops a comprehensive automated pipeline that transforms biplanar EOS images into patient-specific 3D rib cage models through domain adversarial training and geometric reconstruction. To address data limitations, synthetic training datasets are generated from over 750 CT volumes using physics-based Digitally Reconstructed Radiography with multilabel segmentation preservation. A custom domain adversarial neural network architecture employs a twelve-channel output to maintain individual rib continuity while reducing the domain gap between synthetic and clinical images through gradient reversal mechanisms. The final component reconstructs volumetric
models by establishing a mathematical correspondence between orthogonal projections, achieved through centerline extraction, and parametric representation.
The segmentation network achieved Dice scores of 0.95 for coronal and 0.93 for sagittal views on synthetic data, with domain adversarial training producing measurable improvements in generalization to clinical EOS images. The 3D reconstruction component successfully captures patient-specific anatomical characteristics, including the overall shape of the rib cage, individual rib curvature patterns, the spatial relationships between rib pairs, and patient-specific spinal deformations with associated rib adaptations. This work establishes a foundation for extracting comprehensive anatomical information from limited projection data while providing methodological frameworks for medical imaging challenges involving synthetic data generation and cross-domain adaptation.
