Synthetic Data-Driven Deep Learning Approaches for Cone Beam Computed Tomography Applications
Cone beam computed tomography (CBCT) devices are widely used for diagnosis and image-guided interventions, and their mobility and ease of integration with navigation systems give them more clinical flexibility than standard CT. But their smaller detectors cause “truncation,” meaning parts of the anatomy fall outside the field of view and are effectively missing from the reconstruction. This breaks many conventional CBCT algorithms. Deep learning offers a way around this, but training it well runs into two familiar problems: not enough labeled clinical data, and picking the right models for the job. This thesis tackles both, using synthetic data and careful model selection to improve CBCT applications in clinical practice.
A central issue in this thesis is that truncation distorts anything reconstructed outside the field of view, including fiducial markers used for navigation and metallic implants, which throws off both marker detection and metal artifact reduction (MAR). To fix this, the thesis builds deep learning methods for both tasks, backed by several data simulation pipelines designed to work around the shortage of labeled clinical data and the inaccuracies of manual labeling. A task-specific simulation approach improves marker restoration and detection accuracy, while combining 3D projection data with 2D X-ray merging improves MAR performance on real clinical CBCT scans. The thesis also handles both halves of the MAR problem: segmenting the metal and inpainting around it.
On model choice, the thesis draws two key lessons. First, pairing classical algorithms with deep learning beats either alone: the Hough Transform combined with deep learning gives more robust marker detection, and a consistency-check algorithm keeps segmentation results aligned across projections. Second, newer network designs built on Shifted-window (Swin) Transformers outperform conventional convolutional neural networks, showing stronger results in both segmentation and image generation tasks throughout this work.
Put together, these simulation datasets and trained models let the thesis solve clinical imaging problems that conventional methods couldn’t handle, confirmed through evaluation on real clinical data: distorted markers are accurately detected, metal is fully segmented in CBCT scans, and inpainted scans come out cleaner. These methods have already proven their value beyond this thesis too: a Swin Transformer-based network improved CT synthesis from CBCT and MRI in the SynthRAD2023 challenge, and a cross-domain MAR algorithm developed here helped win the 2024 MAR challenge.