Simulating metal ct artefacts for ground truth generation in deep learning.

Sammanfattning: CT scanning stands as one of the most employed imaging techniques used in clinical field. In the presence of metal implants in the field of view (FOV), distortions and noise appear on the 3D image leading to inaccurate bone segmentation, often required for surgery planning or implant design. In this research project, we focused on developing a pipeline to create a rich ground truth dataset, forming the foundation for training a deep-learning bone segmentation algorithm capable of great results even in the presence of metal artefacts. In order to build an extensive artefact-contaminated database, metal-free datasets were collected and bone delineated using Segment 3DPrint automatic bone-segmentation tool. Realistic metal implants were then designed in proper anatomical locations and synthetic artefacts generated using a MATLAB-based algorithm. The effectiveness of the simulator has been tested on real data using an anthropomorphic hand phantom with metal implants inserted and scanned with standard clinical CT parameters. The simulator has proven to successfully mimic physical phenomena such as beam hardening, phantom starvation and noise which are the underlying causes of real metal artefacts. It produces realistic artefacts shapes, even for complex metal configurations. Additional datasets already exhibiting metal artefacts were also added to the database. The simulator was used there only to virtually rescanned those datasets for augmentation reasons. Finally, a training pipeline was imagined using the artefact simulator in parallel to the training process. Data can thus be constantly augmented with new features as the training of the network is running.